- Email: [email protected]
A multi-process system for HEp-2 cells classification based on SVM
ARTICLE IN PRESS
JID: PATREC
[m5G;April 25, 2016;22:24]
Pattern Recognition Letters 0 0 0 (2016) 1–8
Contents lists available at ScienceDirect
Pattern Recognition Letters journal homepage: www.elsevier.com/locate/patrec
A multi-process system for HEp-2 cells classification based on SVM✩ Donato Cascio a,∗, Vincenzo Taormina a, Marco Cipolla a, Salvatore Bruno a, Francesco Fauci b, Giuseppe Raso a a b
Dipartimento di Fisica e Chimica dell, Università degli Studi di Palermo, viale delle Scienze Ed. 18, Palermo 90128, Italy CyclopusCAD ltd, Palermo, Italy
a r t i c l e
i n f o
Article history: Available online xxx Keywords: Hep-2 cells classification Indirect immunofluorescence SVM Accuracy Features reduction
a b s t r a c t This study addresses the classification problem of the HEp-2 cells using indirect immunofluorescence (IIF) image analysis, which can indicate the presence of autoimmune diseases by finding antibodies in the patient serum. Recently, studies have shown that it is possible to identify the cell patterns using IIF image analysis and machine learning techniques. In this paper we describe a system able to classify presegmented immunofluorescence images of HEp-2 cells into six classes. For this study we used the dataset provided for the participation to the “contest on performance evaluation on indirect immunofluorescence image analysis systems”, hosted by the ICPR 2014. This system is based on multiple types of class-process and uses a two-level pyramid to retain some spatial information. We extract a large number (216) of features able to fully characterize the staining pattern of HEp-2 cells. We propose a classification approach based on the one-against-one (OAO) scheme. To do this, an ensemble of 15 support vector machines is used to classify each cell image. Leave-one-specimen-out cross validation method was used for the system optimization. The developed system was evaluated on a blind Hep-2 cells dataset performing a mean class accuracy (MCA) equal to 80.12%. © 2016 Elsevier B.V. All rights reserved.
1. Introduction Autoimmune diseases (AID) are a collection of many complex disorders of unknown aetiology resulting in immune responses to self-antigens and are thought to result from interactions between genetic and environmental factors. Antinuclear antibodies (ANAs) are significant biomarkers in the diagnosis of autoimmune diseases in humans which is done by means of indirect immunofluorescence (IIF) method, and performed by analysing patterns and fluorescence intensity. There are over 80 different AID, and overall they are among the most prevalent diseases in the US, affecting at least 7% of the population. Since most AID are chronic and incurable, from a public health perspective they constitute a major health problem that, besides causing individual suffering, has high societal costs [1]. These diseases can affect people of all ages and both sexes, with a higher frequency in women of child-bearing age. The autoimmune diseases are multi-factorial, and their risk factors are genetic and environmental. The binding of auto-antibodies on HEp-2 cells is revealed by fluorescent antibodies to human immunoglobulin. The fluorescence pattern observed with the mi-
✩ ∗
This paper has been recommended for acceptance by Mario Vento. Corresponding author. Tel.: +39 091 238 99050; fax: +39 091 238 99063. E-mail address: [email protected] (D. Cascio).
croscope (example: homogeneous, speckled, nucleolar, nentromere, etc.) is specific according to the nature of the self antigen and of its location in the cell. Such slides are typically examined by pathologists, however, due to the difficulty of the task, a Computer Aided Design system is desirable [2–4]. However, their pattern identification is often subjective and low standardized. Hence, it is beneficial to develop the automatic stain pattern identification algorithms which can serve as a computer-aided diagnosis (CAD) system. Due to its potential applications, the computer vision based stain patterns identification has attracted much attention [5]. Recently Elgaaied Benammar et al. [6] have optimized and tested a CAD system on HEp-2 images, and preliminary results showed that the CAD, used as second reader, resulted to better performance than Junior immunologists and hence may significantly improve their efficacy; compared with two Junior Immunologists, the CAD system showed higher Intensity accuracy (85.5% versus 66.0% and 66.0%), higher patterns accuracy (79.3% versus 48.0% and 66.2%) and higher mean class accuracy (79.4% versus 56.7% and 64.2%). In this work we focus on building an automated pattern recognition system for the classification of IIF HEp-2 cell images into a set of predefined classes. The system described in this paper participated to the contest related to the special issue of pattern recognition letters titled “pattern recognition techniques for indirect immunofluorescence images analysis”.
http://dx.doi.org/10.1016/j.patrec.2016.03.024 0167-8655/© 2016 Elsevier B.V. All rights reserved.
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
JID: PATREC 2
ARTICLE IN PRESS
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
1.1. Related work This section briefly reviews previous work related to HEp-2 cell image classification in the context of ANA testing. The problem of HEp-2 cell classification attracted major attention among researchers with the benchmarking contests [7,8]. The I3A contest (performance evaluation of indirect immunofluorescence image analysis systems) held in conjunction with ICPR 2014 was the most recent in this series of contests [9]. It received 11 submissions to Task 1 (cell classification) and 7 submissions to Task 2 (specimen classification). In this contest Ensafi et al. [10] proposed a classification method where the SIFT and SURF features are extracted as the input features to learn a dictionary followed by spatial pyramid matching (SPM) to provide the sparse representation of the input cell images. Then a support vector machine (SVM) has been trained to classify the test images. Manivann et al. [11] presented a system to recognize such patterns, at cellular and specimen levels, in images of HEp-2 cells. Ensembles of SVMs were trained to classify cells into six classes based on sparse encoding of texture features with cell pyramids, capturing spatial, multi-scale structure. Mean class accuracies for cell classification obtained on used test data sets were 87.1%. These were the highest achieved in the competition hosted by ICPR2014. At present, there are many mature image recognition and classification algorithms, and bag of features (BOF) algorithm is a popular one among them; BOF algorithm came from the bag of words (BOW) and was originally used for documents matching. In recent years, BOF algorithm is also been applied in the field of HEp-2 cell classification and has achieved results [12]. However, in the BOF algorithm there exists some problem in the application of image recognition and classification such as run is not fast enough and the classification accuracy is not high enough. So it needs to be optimized. Recent research has focused on extracting various morphological features and texture features such as local binary patterns (LBP) have been widely applied for classification of segmented images of HEp-2 cells [14]. Due to poor quality images, utilizing low-level features directly may be inefficient. Nosaka et al. [13] propose co-occurrence of adjacent local binary patterns (CoALBP) to extract textural features. Using linear support vector machine (SVM), their method won the first prize in the contest ICPR2012. Recently [15] a specimen-level image classification system was proposed, in which a specimen-level image descriptor is learned, based on cell-level attributes. In another system, proposed by Soda et al. [16], each specimen image is classified based on the strength of the recognition of the cells in that specimen image. Han et al. [17] have employed a parametric probability process to model local image patches (textons: microstructures in the cell image) and extract the higher-order statistics of the model parameters for the image description. They have used a simple linear support vector machine for cell pattern identification. Cascio et al. [18] proposed a classification approach based on two steps: the first step follows the one-against-all (OAA) scheme, while the second step follows the one-against- one (OAO) scheme. To do this, they have implemented 21 KNN classifiers: 6 OAA and 15 OAO. Perner et al. [19] presented an early attempt on developing an automated HEp-2 cell classification system. Cell regions were represented by a set of basic features extracted from binary images obtained at multiple grey level thresholds. Those features were then classified into six categories by a decision tree algorithm. Theodorakopoulos et al. [20] combined a set of local features, including LBP and rotation-invariant SIFT, with vectors of locally aggregated descriptors (VLAD).
Fig. 1. IIF images with different staining patterns (from left to right and from up to down: homogeneous, speckled, nucleolar, centromere, Golgi and nuclear membrane).
1.2. Our contributions In this paper we describe a system for the classification of pre-segmented immunofluorescence images of HEp-2 cells into six classes: homogeneous, speckled, nucleolar, centromere, Golgi, and nuclear membrane. Fig. 1 shows examples of each class. This system is based on multiple types of class-process which uses a twolevel pyramid to retain some spatial information. The basic intuition behind our approach is that, instead of using a single process to discriminate each class from all other classes, it is better to combine a set of different and complementary processes. For this purpose, an intensive analysis of preprocessing types has been conducted and for each process it has been identified the preprocessing method giving the best performance in terms of final class accuracy. That analysis was carried out by distinguishing between the mask cells and the adjacent circular crown, and 30 different best preprocessing were obtained. We extracted a large number of features (216), and for each classification process a phase of features reduction, based on Fisher criterion, was performed in order to select the best features, able to fully characterize the staining pattern of HEp-2 cells. We discuss the details of our approach in the following sections. As final consideration, we want to highlight that the method here presented has the advantage of allowing an easy extension to the problem of aided diagnosis, and hence to the development of a CAD for the support to diagnosis, which is the ultimate goal of this research [6], this is possible by including an automatic segmentation phase in the system, also differentiated (given the enormous variety of ways in which the pattern may occur, this is probably the only way to achieve an efficient system). 2. Materials and methods In several multiclass classification problems, it is preferable to use a number of classifiers equal to the number of classes and each classifier is trained in order to discriminate a class from all the others (binary approach) [21]. In this work, the classification stage is differentiated by using 15 one-against-one classifiers for six classes of staining patterns. In addition, the preprocessing and feature extraction steps are differentiated too. Therefore, we adopt a non-standard pipeline for supervised image classification. The generic image is simultaneously processed by 15 processes obtaining 15 separate outputs that represent how the cell resembles each one of the 6 classes analysed in this work. This system
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
JID: PATREC
ARTICLE IN PRESS
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
3
for each of the 30 OAO developed classifiers, obtained after an intensive and exhaustive analysis of the combinations of functions, taken two by two, reported in Table 1. The description of methods for selection of optimal preprocessing algorithms and the flow chart of the optimization method used for each (fifteen) classifiers one-against-one developed is shown in Section 2.4. As it can be easily seen from Table 2, the preprocessing algorithm obtained with contrast normalization and dilation has been associated with most types of OAO classifiers. 2.2. Features extraction and dimensionality reduction
Fig. 2. Pipeline of developed method.
uses a two-level pyramid to retain some spatial information. In fact, in addition to the segmented mask regions, the information is extracted from circular regions adjacent to the cells. For each image cell, the mask used for the segmentation was dilated (using a 7 × 7 structuring element) to obtain the adjacent circular crowncontaining information on the image area next to the cell. For each process a preprocessing phase is performed and a set of specific features are extracted both from these cells and from the adjacent circular crown. Finally a RBF support vector machine (SVM) classifier was employed to classify each image into one of two specific class patterns to which the specific classification process is addressed (binary approach, one-against-one). Fig. 2 shows the flow of operations adopted in this work. The following sections describe the proposed method in some more detail. The choice of methods, features and parameters was performed automatically, using the mean class accuracy (MCA) as a figure of merit. The parameters involved are tuned using a leave-onespecimen-out cross validation scheme. The main benefit of this pipeline is that it has a good explanatory power, as its principle is simple to explain.
The cells patterns analysis clearly shows that differences between classes are mainly based on the presence and distribution of bright/dark structures: numbers, sizes (areas), intensity and colours. It seems natural to use features providing an analysis of texture and more specifically features able to deal with bright/dark speckle-like structure description. Different staining patterns can be characterized by a limited set of attributes describing the spatial relationships between pixels values and the main image variations occurring in each cell type [22]; this information is generally obtained by mean of textural analysis techniques. These techniques can be grouped into two major categories: (i) statistical methods, describing the distribution of grey-levels in the image, and (ii) frequency domain measurements of image variations [23–25]. To achieve the objective of robust classification, we combined several discriminative visual features known to be effective for cell classifications with a robust and scalable multi-class boosting. All classifiers use a large number of extracted features, able to fully characterize the HEp-2 cells. In more details, four quantization levels [19, 26] were analysed: 8 bit, 7 bit, 6 bit and 5 bit of histogram intensity (respectively related to 256, 128, 64 and 32 Gy levels). For each quantization level the following 27 features have been extracted: •
•
•
2.1. Preprocessing The preprocessing phase in a pattern recognition task as that here described, where the type of objects to be classified is quite heterogeneous, is very important, and surely affects the next phase of features extraction, and hence the classification itself. This is the reason why, as above anticipated, in this work we decided to differentiate the preprocessing phase for each type of pattern to be recognized. For this purpose, starting from a set of functions, analyzed and reported in Table 1, many combinations of different types of preprocessing have been analysed and for each class it has been identified the preprocessing method giving the best performance in terms of final class accuracy. For many functions reported in Table 1, several different configurations have been evaluated, using several parameters (e.g. for the Gaussian filter the convolution mask, sized 3 × 3, 5 × 5 and 7 × 7 have been analysed), leading to a total of 37 analysed functions. Table 2 shows the details of combinations of preprocessing and their use in one-against-one classification processes. The analysis was conducted by distinguishing between the mask cells and the circular crown, for which, in fact, 30 different best preprocessing methods were obtained. In Table 2 the best results were reported
•
Intensity based features (9): mean value, standard deviation, ratio of the standard deviation to the mean value, entropy, moment of inertia, skewness, kurtosis, entropy of the contours gradient [27–29]. Geometry-based features (8): mean radius, standard deviation of radius, maximum radius, ratio of the standard deviation to the mean value, circularity, anisotropy, fractal index, eccentricity [30]. Shape-morphological-based features (8): area, perimeter, convex area, convex deficiency, solidity, compactness, roundness, Euler’s number. Descriptors-based features (2): entropy of HOG (histogram of oriented gradients), entropy of HAG (histogram of amplitude gradients).
From each cell image, all the features were densely extracted from image patches and from the circular crown obtained by mask dilation. To summarize, the 27 features above mentioned, being extracted using 4 quantization levels from image patches and from the circular crown, led to get a total analysis over 216 features (27 × 4×2 = 216). Reducing the dimensionality of the data by selecting a subset of the original variables may be advantageous to simplify the classification problem from different points of view. As the dimensionality of the data increases, many types of data analysis and classification problems become significantly harder. Sometimes the data also becomes increasingly sparse in the feature space. This can lead to big problems for both supervised and unsupervised learning. The main idea of feature subset selection is to remove redundant or irrelevant features from the dataset as they can negatively influence the classification accuracy, the clustering quality and lead to an unnecessary increase of computational cost. The
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC 4
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8 Table 1 List of functions analysed for preprocessing. Abbreviation
Function description
Reference
Nt Cn
Nothing Contrast normalization: was obtained by linearly remapping the intensity values so that 1% of data is saturated at low and high intensities. Equalization CLAHE (contrast limited adaptive histogram equalization) Median filter Gaussian filter Gaussian-enhancement: Dilatation Erosion Nagao Bilateral Opening FAS (filter alternate segmental)
http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://www.pkuwwt.tk/ofeli/doc/index.html http://docs.opencv.org http://docs.opencv.org http://www.pkuwwt.tk/ofeli/doc/index.html
Eq Ch Md Gs Ge Di Er Ng Bl Op Fs
Table 2 List of preprocessing used in OAO classifiers. Preprocessing for mask cell
Preprocessing for circular crown
N
Function 1
Function 2
Function 1
Function 2
Used for classification:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Cn Di Di Cn Cn Cn Cn Cn Cn Cn Cn Bl Cn Cn Op
Ng Cn Cn Bl Bl Bl Di Bl Bl Ge Ng Cn Bl Gs Cn
Di Di Di Eq Eq Di Cn Eq Cn Di Di Di Cn Di Cn
Cn Cn Cn Op Di Cn Fs Op Di Ch Cn Cn Gs Cn Gs
Homogeneous versus speckled Homogeneous versus nucleolar Homogeneous versus centromere Homogeneous versus Golgi Homogeneous versus nuclear membrane Speckled versus nucleolar Speckled versus centromere Speckled versus golgi Speckled versus nuclear membrane Nucleolar versus centromere Nucleolar versus Golgi Nucleolar versus nuclear membrane Centromere versus Golgi Centromere versus nuclear membrane Golgi versus nuclear membrane
advantage of identifying a subset of features is that no information about the importance of single features is lost. Filter methods perform feature selection in two steps. In the first step, the filter method assesses each feature individually for its potential in discriminating among classes in the data. In the second step, features falling beyond some threshold criterion are eliminated, and the smaller set of remaining features is used. This score-and-filter approach has been used in many recent publications, due to its relative simplicity [31,32]. Scoring methods generally focus on measuring the differences between distributions of features. The resulting score is intended to reject the quality of each feature in terms of its discriminative power. Many scoring criteria exist, in this work we have adopted the criterion of statistical selection of Fischer (Eq. (1)). By this criterion, the quality of the ith feature is expressed in terms of the difference among the empirical means of two distributions, normalized by the sum of their variances.
V (i ) =
(μ1 (i ) − μ2 (i ) )2 σ12 + σ22
(1)
In our work, μ1 (i) and μ2 (i) represent the average values of ith feature respectively on pattern 1 and on pattern 2 (OAO classification), while σ 1 and σ 2 represent the corresponding standard deviations. For each pair of patterns to be classified, the Fisher score is therefore evaluated on each extracted feature, and only the 15 highest values obtained from the cells and the 15 highest values obtained from their circular crowns (30 features out of 216) were selected. All the developed classifiers have 30 input, of course not necessarily the same.
2.3. Classification The classification is a crucial task for this problem and requires a training phase computationally expensive. The conventional way to deal with a multi-class problem is to decompose an M-class problem into a series of two-class problems. The two approaches commonly used are the One-Against-One (OAO) and One-AgainstAll (OAA) techniques. The OAA approach [33,34] represents the earliest and most commonly used multiclass approach and involves the division of an M class dataset into M two-class cases; OAA uses M classifiers where the ith classifier separates class i from all the remaining classes. In order to assign a class to a new sample, it is necessary to evaluate the output for all binary classifiers and, usually, choosing the pattern relative to the classifier that returned the highest classification value. The problem with the participation to the decision is assumed to be equally reliable, which is rarely the case. The OAO approach on the other hand involves constructing a classifier for each pair of classes resulting in M(M–1)/2 classifier; the OAO approach is more computationally intensive. When applied to a test point, each classification gives one vote to the winning class and the point is labeled with the class having most votes. One drawback of one-against-one method, however, arises when the results from the multiple classifiers are counted for the final decision without considering the competence of the classifiers. In this paper we propose an approach based on OAO approach. For each pair of class patterns one classifier is implemented (e.g. homogeneous versus speckled), for a total of 15 classifiers (see Fig. 2). The final decision is obtained counting the obtained votes by each class patterns and choosing the one with the
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
5
with the higher discriminating power, and this is done by selecting, according to the Fisher criterion, the best 15 features for the masks and the best 15 features for circular crowns. For each combination of preprocessing functions is evaluated both the value DM given by the sum of the 15 best values of Fisher for the masks, and the sum of the DC of the 15 best values of Fisher for the circular crowns.
DM =
15
Vmask (i )
(3)
Vcrown (i )
(4)
i=1
DC =
15 i=1
The selection of the best preprocessing functions for each process was carried out in order to maximize the sum DM + DC ; for each binary classification, the configuration having maximum DM is stored and at the end is coupled to the one with maximum DC . The configuration of functions and features thus obtained, for the masks and for the circular crowns, is used for the subsequent step of identification of the parameters of the classifier. In the end, for each binary classifier, it is identified a set of preprocessing functions, a set of features, and an SVM with the relevant parameters. The SVMs training was carried out to maximize the MCA, using the leave-one-specimen-out cross validation cross procedure. In the end the optimization method leads to a multi-process system as represented in the schema of Fig. 2. 2.5. Leave-one-specimen-out
Fig. 3. Flow chart of the optimization method used for each developed (fifteen) one-against-one classifier.
most votes. SVMs (Support Vector Machines) are a useful technique for data classification, because, unlike many classifiers in which it is easy to overfit the data, SVM provides a better generalization bound. In this work it was decided to use SVM classifiers. The (Gaussian) RBF (Radial Basic Function) kernel is a reasonable first choice. This kernel nonlinearly maps samples into a higher dimensional space, so that, unlike the linear kernel, can handle the case when the relation between class labels and attributes is nonlinear. Furthermore, the linear kernel is a special case of RBF [35,36] since the linear kernel with a penalty parameter C has the same performance as the RBF kernel with some parameters (C, γ ). In addition, the sigmoid kernel behaves like RBF for certain parameters [35,36]. In this work we adopted the following function kernel:
2
k(x, y ) = exp −γ xi − x j ,
γ >0
(2)
2.4. Method for functions and parameters selection In the present work an intensive analysis of methods and parameters was performed, in order to analyze and classify in different ways the patterns sought. This system indeed, for each binary classification, identifies: a set of preproccesing functions (4 functions), a set of features (30 features) and a set of parameters for the SVM classifier (C and γ ). In Fig. 3 it is shown the flow chart of the optimization method used for each of the 15 implemented processes. As it can be seen from Fig. 3, any two preprocessing functions between those implemented (shown in Table 1) are combined in a comprehensive manner and the related preprocessing is performed. The images thus obtained are used for the extraction of features, both for the masks and for the circular adjacent regions. The next phase of features has the aim of identifying the features
In this work the leave-one-specimen-out cell cross validation technique [11,37] have been used to exploit the highest possible number of patterns during the training phase, without invalidating the results. The method consists in leaving out all cells belonging to the same specimen, rather than leaving out a single cell for the construction of the training set; cells of the same specimen are similar in terms of the average intensity and contrast, and introduce bias. The leave-one-specimen-out strategy is an effective way to evaluate the algorithm when the test data set is not available. 2.6. Statistics calculation The Accuracy and the mean class accuracy are adopted in this work as measures of the performance [38]. Let CCRk be the correct classification rate for class k determined as follows:
CCRk =
Tk Nk
(5)
where Tk is the number of correct identifications of class k, while Nk is the total number of elements of class k. The accuracy is defined by:
Accuracy =
k
CCR · N k k k Nk
(6)
The MCA is determined by:
MCA =
1 CCRk k k
(7)
3. Results 3.1. Dataset The development of a CAD system is intimately linked to collection of a dataset of selected images [39]. For this study we used
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC 6
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8 Table 3 Best values for C and γ parameters.
Fig. 4. Distribution of training set patterns.
the dataset provided for the participation to the “contest on performance evaluation on indirect immunofluorescence image analysis systems”, hosted by the 22nd International Conference on Pattern Recognition (ICPR 2014); the dataset can be downloaded after proper registration on site http://i3a2014.unisa.it/. The participants are required to face exactly the same task of the ICIP 2013 Contest by using the same training and testing dataset [40]. The goal is the design and implementation of an IIF pattern recognition system able to classify the cells belonging to HEp-2 images in one of the following pattern classes: homogeneous, speckled, nucleolar, centromere, Golgi and nuclear membrane. The dataset has been collected between 2011 and 2013 at Sullivan Nicolaides Pathology laboratory, Australia. It utilizes 419 patient positive sera, which were prepared on the 18-well slide of HEP-20 0 0 IIF assay from Immuno Concepts N.A. Ltd. with screening 1:80 dilution. The specimens were then automatically photographed using a monochrome high dynamic range cooled microscopy camera which was fitted on a microscope with a plan-Apochromat 20x/0.8 objective lens and an LED illumination source. Approximately 100– 200 cell images were extracted from each patient serum. In total there were 13,596 cell images used for training. The labelling process involved at least two scientists who read each patient specimen under a microscope. A third expert’s opinion was sought to adjudicate any discrepancy between the two opinions. We used each specimen label for the ground-truth of cells extracted from it. Furthermore, all the labels were validated by using secondary tests such as ENA and anti-ds-DNA in order to confirm the presence and/absence of specific patterns. Each cell image contained in the database is annotated with the following information: cell pattern (one of the patterns above defined), cell intensity, cell mask, ID of the image which the cell belongs to. A characterization of the dataset is reported in Fig. 4. 3.2. SVM parameters There are two parameters for an RBF kernel: C and γ . It is not known beforehand which C and γ are best for a given problem; consequently some kind of model selection (parameter search) must be done. The goal is to identify good (C; γ ) so that the classifier can accurately predict unknown data (i.e. testing data). Note that it may not be useful to achieve high training accuracy (i.e. a classier which accurately predicts training data whose class labels are indeed known). A common strategy is to separate the dataset into two parts, of which one is considered unknown. The prediction accuracy obtained from the “unknown" set more precisely reflects the performance on classifying an independent dataset. We have used a “grid-search" on C and γ using leave-one-specimenout validation. Various pairs of (C, γ ) values are tried and the one
Class
C
γ
Homogeneous versus speckled Homogeneous versus nucleolar Homogeneous versus centromere Homogeneous versus Golgi Homogeneous versus nuclear membrane Speckled versus nucleolar Speckled versus centromere Speckled versus Golgi Speckled versus nuclear membrane Nucleolar versus centromere Nucleolar versus Golgi Nucleolar versus nuclear membrane Centromere versus Golgi centromere versus nuclear membrane Golgi versus nuclear membrane
8 8 1024 4 32 4 4 4 4 4 4 64 4 32 4
1.0 0.5 7.8 × 10−3 7.8 × 10−3 1.3 × 10−1 7.8 × 10−3 3.1 × 10−2 1.6 × 10−2 1.3 × 10−1 7.8 × 10−3 7.8 × 10−3 1.6 × 10−2 1.0 1.6 × 10−2 3.1 × 10−2
Table 4 Per-class accuracy. Class
Accuracy
Centromere Golgi Homogeneous Nucleolar Nuclear membrane Speckled
87.09% 88.54% 93.70% 89.26% 90.24% 88.70%
with the best accuracy is taken. We found that trying exponentially growing sequences of C and γ is a practical method to identify good parameters [35,36]. In this paper, the probed values for C and γ were:
C = 2−5 , 2−4 , . . . , 210 ; γ = 2−10 , 2−9 , . . . , 22 For each pair of class pattern the analyzed grid had size equal to 16 × 13, i.e. 208 points. In Table 3 are reports the best values of parameters C and γ obtained for each pair of class pattern. 3.3. Experimental results The proposed approach has been trained on the dataset provided by the ICPR 2014 contest for task 1. Task 1 is delegated to cell level classification. In particular, it consists of 13,596 images for the training phase coming from 419 patient positive sera and categorized by trained physicians into six patterns: centromere, homogeneous, nucleolar, speckled (coarse and fine), nuclear membrane and Golgi. The classification results obtained in our experiments have been summarized in Table 4, and organized by staining pattern class; where, for each of them, we show the obtained accuracy. The estimated mean class accuracy for the proposed method is 90.25%, with a maximum per-class accuracy of 93.7% for homogeneous patterns. From the table we can notice that the poorest classification accuracy is achieved by centromere pattern class. For some patterns, such as the Golgi, the improvement in classification performance was significant, due to use of the information contained in circular crowns. Furthermore, the software implementation of the method presented here has been submitted to the group owning the database used in ICPR2014 contest [41], and this allowed to evaluate the system’s performance on the same “test set” used in task 1 of ICPR2014 contest; this dataset was not used in training phase. The performances results are reported in the confusion matrix presented in the Table 5. The mean class accuracy (MCA) obtained is equal to 80.12%, with a maximum per-class accuracy of 93.7% for homogeneous
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8 Table 5 Confusion matrix of our method.
Centr Golgi Homog Nucle NucMe Speckl
Centr
Golgi
Homog
Nucle
NucMe
Speckl
93.7% 0.03% 0.02% 1.6% 0.2% 15.0%
0,2% 76.4% 0.7% 0.7% 2.3% 0.5%
0.7% 2.0% 73.8% 2.2% 5.2% 13.3%
2.3% 8.1% 4.4% 85.1% 4.3% 2.2%
0.3% 11.3% 13.1% 2.6% 86.7% 3.9%
2.9% 2.2% 8.0% 7.8% 1.5% 65.2%
Fig. 5. Mean accuracy obtained by all the methods on the test set ICPR2014.
patterns, and a minimum per-class accuracy of 65.2% for speckled patterns. In order to have a comparison with other developed methods presented in other works, a graph (Fig. 5) showing the performance obtained, on the same dataset, by other research groups at the ICPR 2014 context, is following reported. As it can be easily deduced by Fig. 5, the method presented in this work is definitely comparable, in terms of performance, with the results obtained by other research groups in the recent works. 3.4. Running time The average running time to analyse a cell image and its mask is 0.8 s, about 2 min per IIF image (using a 3.4 GHz Intel i7 CPU). The training time, using Fisher analysis, for each classifier, takes about 4 h. The tuning of SVMs takes about 45 min for each classifier. Note that it is possible to reduce the total running time by parallelizing each of the fifteen independent processes. 4. Discussion and conclusions In this paper we proposed an approach for the automatic classification of staining patterns in HEp-2 cell IIF images, which is critical for the diagnosis of autoimmune diseases. We adopted a non-standard pipeline for supervised image classification that provides a diversified process for all types of class-patterns to classify. For this purpose, some different types of preprocessing algorithms have been analysed and for each process it has been identified the preprocessing method giving the best performance in terms of final class accuracy. We extracted a large number of features (216), able to fully characterize the staining pattern of HEp-2 cells. Then, for each class, a sub-set of 30 features has been selected. Furthermore, we proposed a classification-approach based on OAO classification rule. Our classification system is thus composed by 15 SVM classifiers, with an approach one-against-one. Leave-one-specimenout cross validation method was used for the optimization of the system. We evaluated the method performances on the HEp-2 cells dataset (manually segmented and annotated). Although being the
early results of our methodology for such a challenging application, performances are really satisfactory (mean class accuracy of 80.12%). The choice to combine the OAO approach, therefore, produces an improvement, even if not notable, of the mean class accuracy. The significant difference between the results, presented in the Table 3, obtained with leave-one-out method and those of the confusion matrix (Table 4) is certainly to ascribe to the presence of bias. The present method has certainly the advantage that can be easily extended to research for additional autoantibody patterns having simply to add, in this case, the classification processes related to possible new patterns to search. In addition, the intensive analysis conducted regarding preprocessing functions and parameters can be used in order to get information for a study based on LDA or similar methods. The used method of reduction of dimensionality, compared to other criteria, highly reduces the computational load and, in that sense, allowed to perform the intensive analysis of functions and parameters carried out for each of the 15 implemented classification processes. The major limitation of the method is certainly linked to the lack of correlation analysis between the features (that, on the contrary, the LDA is able to do); two totally related features may be both selected by the Fisher method. The performance of the proposed method can be easily improved. In fact, the method of coupling the results of the 15 classifiers obtained as simple count of votes, is a simplistic solution, probably the inclusion of a final classifier (such as K-NN) could have better enhanced the conducted analysis. Further analysis is being performed to evaluate the robustness and reliability of other features selection methods. This includes the evaluation of exhaustive or almost-exhaustive selection approaches, by focusing on the feature space proposed in this work.
References [1] National Institutes of Health, The autoimmune disease coordinating committee, Progress Autoimmune Dis. Res. (2005) 1–146. [2] X. Xiang, L. Feng, C. Ng, L.K. Pang, Staining pattern classification of ANA-IIF based on SIFT features, J. Med. Imag. Health Inf. 2 (4) (2012) 419–424. [3] S. Ciatto, et al., Computer assisted diagnosis (CAD) in mammography. Comparison of diagnostic accuracy of a new algorithm (Cyclopus®, Medicad) with two commercial systems, Radiol. Med. 114 (4) (2009) 626–635. [4] D. Cascio, F. Fauci, M. Iacomi, G. Raso, R. Magro, D. Castrogiovanni, G. Filosto, R. Ienzi, M. Vasile, Computer-aided diagnosis in digital mammography: comparison of two commercial systems, Imag. Med. 6 (1) (2014) 13–30. [5] A. Rigon, F. Buzzulini, P. Soda, L. Onofri, L. Arcarese, G. Iannello, A. Afeltra, Novel opportunities in automated classification of antinuclear antibodies on HEp-2 cells, Autoimmunity Rev. 10 (10) (2011) 647–652. [6] A. Elgaaied Benammar, et al., Computer-assisted classification patterns in autoimmune diagnostics: the A.I.D.A. Project, Biomed. Res. Int. 2016 (2016) 1–9. [7] P. Foggia, G. Percannella, P. Soda, M. Vento, Benchmarking hep-2 cells classification methods, IEEE Trans. Med. Imag. 32 (10) (2013) 1878–1889. [8] P. Foggia, G. Percannella, A. Saggese, M. Vento, Pattern recognition in stained HEp-2 cells: Where are we now? Pattern Recognit. 47 (7) (2014) 2305–2314. [9] [9] Lovell, B.C., Percannella, G., Vento, M., Wiliem, A., 2014, Performance Evaluation of Indirect Immunofluorescence Image Analysis Systems, Technical Report, ICPR Workshop. [10] S. Ensafi, S. Lu, A.A. Kassim, C.L. Tan, in: A bag of words based approach for classification of hep-2 cell images. Pattern recognition techniques for indirect immunofluorescence images (I3A), 2014 1st Workshop, 2014, pp. 29–32. [11] S. Manivannan, W. Li, S. Akbar, R. Wang, J. Zhang, S. McKenna, An automated pattern recognition system for classifying indirect immunofluorescence images of HEp-2 cells and specimens, Pattern Recognit. 51 (2016) 12–26. [12] L. Shen, J. Lin, S. Wu, S. Yu, HEp-2 image classification using intensity order pooling based features and bag of words, Pattern Recognit. 47 (7) (2014) 2419–2427. [13] R. Nosaka, K. Fukui, HEp-2 cell classification using rotation invariant co-occurrence among local binary patterns, Pattern Recognit. 47 (7) (2014) 2428–2436. [14] A.B.L Larsen, J.S. Vestergaard, R. Larsen, HEp-2 cell classification using shape index histograms with donut-shaped spatial pooling, IEEE Trans. Med. Imag. 33 (7) (2014) 1573–1580. [15] B.C.L.A. Wiliem, P. Hobson, Discovering discriminative cell attributes for hep-2 specimen image classification, in: IEEE Winter Conference on Applications of Computer Vision, Article number 6836071, 2014, pp. 423–430.
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
7
JID: PATREC 8
ARTICLE IN PRESS
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
[16] P. Soda, G. Iannello, Aggregation of classifiers for staining pattern recognition in antinuclear autoantibodies analysis, IEEE Trans. Inf. Technol. Biomed. 13 (3) (2009) 322–329. [17] X.-H. Han, J. Wang, G. Xu, Y.-W. Chen, High-order statistics of microtexton for hep-2 staining pattern classification, IEEE Trans. Biomed. Eng. 61 (2014) 2223–2234. [18] D. Cascio, V. Taormina, M. Cipolla, F. Fauci, M. Vasile, G. Raso, HEp-2 cell classification with heterogeneous classes-processes based on K-nearest neighbours, in: IEEE Computer Society – Proceedings of 1st Workshop on Pattern Recognition Techniques for Indirect Immunofluorescence Images, 2014, pp. 10–15. [19] P. Perner, H. Perner, B. Muller, Mining knowledge for Hep-2 cell image classification, Artif. Intell. Med. 26 (1–2) (2002) 161–173. [20] I. Theodorakopoulos, D. Kastaniotis, G. Economou, S. Fotopoulos, HEp-2 cells classification using morphological features and a bundle of local gradient descriptors, in: 1st Workshop on Pattern Recognition Techniques for Indirect Immunofluorescence Images (I3A), 2014, pp. 33–36. [21] J Furnkranz, Round Robin classification, J. Mach. Learn. Res. 2 (4) (2002) 721–747. [22] A. Sriram, S. Ensafi, S. Roohi, A.A. Kassim, Classification of human epithelial type-2 cells using hierarchical segregation, in: 13th International Conference on Control Automation Robotics Vision (ICARCV), 2014, 2014, pp. 323–328. [23] G.L. Masala, et al., Comparative study of feature classification methods for mass lesion recognition in digitized mammograms, Nuovo Cimento C 30 (3) (2007) 305–316. [24] L. Vivona, et al., Fuzzy technique for microcalcifications clustering in digital mammograms, BMC Med. Imag. 14 (1) (2014) 1–14. [25] G.L. Masala, B. Golosio, et al., Classifiers trained on dissimilarity representation of medical pattern: A comparative study, Nuovo Cimento C 28 (6) (2005) 905–912. [26] M. Iacomi, et al., Mammographic images segmentation based on chaotic map clustering algorithm, BMC Med. Imag. 14 (1) (2014) 1–11. [27] L. Vivona, et al., A fuzzy logic C-means clustering algorithm to enhance microcalcifications clusters in digital mammograms, in: Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) IEEE, Article number 6152551, 2011, pp. 3048–3050.
[28] F. Fauci, et al., A Fourier based algorithm for microcalcifications enhancement in mammographic images, in: IEEE Nuclear Science Symposium and Medical Imaging Conference, Article number 4774254, 2012, pp. 4388–4391. [29] D. Cascio, et al., Mammogram segmentation by contour searching and mass lesions classification with neural network, IEEE Trans. Nucl. Sci. 53 (5) (2006) 2827–2833. [30] D. Cascio, R. Magro, F. Fauci, M. Iacomi, G. Raso, Automatic detection of lung nodules in CT datasets based on stable 3D mass-spring models, Comput. Biol. Med. 42 (11) (2012) 1098–1109. [31] P. Duda, D.G. Stork, Pattern Classification, Wiley-Interscience Publication, 2001. [32] Q. Gu, Z. Li, J. Han, Generalized Fisher score for feature selection, in: Proceedings of the 27th Conference on Uncertainty in Artificial Intelligence, 2011, pp. 266–273. [33] B. Liu, Z. Hao, E.C.C. Tsang, Nesting one-against-one algorithm based on SVMs for pattern classification, IEEE Trans. Neural Netw. 19 (12) (2008) 2044–2052. [34] R. Debnath, N. Takahide, H. Takahashi, A decision based one-against-one method for multi-class support vector machine, Pattern Anal. Appl. 7 (2) (2004) 164–175. [35] C.-C. Chang, C.-J. Lin, LIBSVM: a Library for Support Vector Machines, ACM Trans. Intell. Syst. Technol. 2 (3) (2011) 1–27, Article 27. [36] C.-W. Hsu, C.-C. Chang, C.-J. Lin, A Practical Guide to Support Vector Classification, Department of Computer Science National Taiwan University, Taipei, Taiwan, 2003. [37] A.B.L. Larsen, J.S. Vestergaard, R. Larsen, HEp-2 cell classification using shape index histograms with donut-shaped spatial pooling, IEEE Trans. Med. Imag. 33 (7) (2014) 1573–1580. [38] A. Rigon, F. Buzzulini, P. Soda, G. Iannello, Afeltra, A Novel opportunities in automated classification of antinuclear antibodies on HEp-2 cells, Autoimmunity Rev. 10 (10) (2014) 647–652. [39] S. Tangaro, et al., MAGIC-5: an Italian mammographic database of digitized images for research, Radiol. Med. 113 (4) (2008) 477–485. [40] P. Hobson, B. Lovell, G. Percannella, M. Vento, A. Wiliem, Benchmarking human epithelial type 2 interphase cells classification methods on a very large dataset, Artif. Intell. Med. 65 (3) (2015) 239–250. [41] Techniques for Indirect Immunofluorescence Images (I3A), 2014 ICPR2014 contest http://nerone.diem.unisa.it/i3a/wp-content/uploads/ICPR2014_Report.pdf
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
JID: PATREC
[m5G;April 25, 2016;22:24]
Pattern Recognition Letters 0 0 0 (2016) 1–8
Contents lists available at ScienceDirect
Pattern Recognition Letters journal homepage: www.elsevier.com/locate/patrec
A multi-process system for HEp-2 cells classification based on SVM✩ Donato Cascio a,∗, Vincenzo Taormina a, Marco Cipolla a, Salvatore Bruno a, Francesco Fauci b, Giuseppe Raso a a b
Dipartimento di Fisica e Chimica dell, Università degli Studi di Palermo, viale delle Scienze Ed. 18, Palermo 90128, Italy CyclopusCAD ltd, Palermo, Italy
a r t i c l e
i n f o
Article history: Available online xxx Keywords: Hep-2 cells classification Indirect immunofluorescence SVM Accuracy Features reduction
a b s t r a c t This study addresses the classification problem of the HEp-2 cells using indirect immunofluorescence (IIF) image analysis, which can indicate the presence of autoimmune diseases by finding antibodies in the patient serum. Recently, studies have shown that it is possible to identify the cell patterns using IIF image analysis and machine learning techniques. In this paper we describe a system able to classify presegmented immunofluorescence images of HEp-2 cells into six classes. For this study we used the dataset provided for the participation to the “contest on performance evaluation on indirect immunofluorescence image analysis systems”, hosted by the ICPR 2014. This system is based on multiple types of class-process and uses a two-level pyramid to retain some spatial information. We extract a large number (216) of features able to fully characterize the staining pattern of HEp-2 cells. We propose a classification approach based on the one-against-one (OAO) scheme. To do this, an ensemble of 15 support vector machines is used to classify each cell image. Leave-one-specimen-out cross validation method was used for the system optimization. The developed system was evaluated on a blind Hep-2 cells dataset performing a mean class accuracy (MCA) equal to 80.12%. © 2016 Elsevier B.V. All rights reserved.
1. Introduction Autoimmune diseases (AID) are a collection of many complex disorders of unknown aetiology resulting in immune responses to self-antigens and are thought to result from interactions between genetic and environmental factors. Antinuclear antibodies (ANAs) are significant biomarkers in the diagnosis of autoimmune diseases in humans which is done by means of indirect immunofluorescence (IIF) method, and performed by analysing patterns and fluorescence intensity. There are over 80 different AID, and overall they are among the most prevalent diseases in the US, affecting at least 7% of the population. Since most AID are chronic and incurable, from a public health perspective they constitute a major health problem that, besides causing individual suffering, has high societal costs [1]. These diseases can affect people of all ages and both sexes, with a higher frequency in women of child-bearing age. The autoimmune diseases are multi-factorial, and their risk factors are genetic and environmental. The binding of auto-antibodies on HEp-2 cells is revealed by fluorescent antibodies to human immunoglobulin. The fluorescence pattern observed with the mi-
✩ ∗
This paper has been recommended for acceptance by Mario Vento. Corresponding author. Tel.: +39 091 238 99050; fax: +39 091 238 99063. E-mail address: [email protected] (D. Cascio).
croscope (example: homogeneous, speckled, nucleolar, nentromere, etc.) is specific according to the nature of the self antigen and of its location in the cell. Such slides are typically examined by pathologists, however, due to the difficulty of the task, a Computer Aided Design system is desirable [2–4]. However, their pattern identification is often subjective and low standardized. Hence, it is beneficial to develop the automatic stain pattern identification algorithms which can serve as a computer-aided diagnosis (CAD) system. Due to its potential applications, the computer vision based stain patterns identification has attracted much attention [5]. Recently Elgaaied Benammar et al. [6] have optimized and tested a CAD system on HEp-2 images, and preliminary results showed that the CAD, used as second reader, resulted to better performance than Junior immunologists and hence may significantly improve their efficacy; compared with two Junior Immunologists, the CAD system showed higher Intensity accuracy (85.5% versus 66.0% and 66.0%), higher patterns accuracy (79.3% versus 48.0% and 66.2%) and higher mean class accuracy (79.4% versus 56.7% and 64.2%). In this work we focus on building an automated pattern recognition system for the classification of IIF HEp-2 cell images into a set of predefined classes. The system described in this paper participated to the contest related to the special issue of pattern recognition letters titled “pattern recognition techniques for indirect immunofluorescence images analysis”.
http://dx.doi.org/10.1016/j.patrec.2016.03.024 0167-8655/© 2016 Elsevier B.V. All rights reserved.
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
JID: PATREC 2
ARTICLE IN PRESS
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
1.1. Related work This section briefly reviews previous work related to HEp-2 cell image classification in the context of ANA testing. The problem of HEp-2 cell classification attracted major attention among researchers with the benchmarking contests [7,8]. The I3A contest (performance evaluation of indirect immunofluorescence image analysis systems) held in conjunction with ICPR 2014 was the most recent in this series of contests [9]. It received 11 submissions to Task 1 (cell classification) and 7 submissions to Task 2 (specimen classification). In this contest Ensafi et al. [10] proposed a classification method where the SIFT and SURF features are extracted as the input features to learn a dictionary followed by spatial pyramid matching (SPM) to provide the sparse representation of the input cell images. Then a support vector machine (SVM) has been trained to classify the test images. Manivann et al. [11] presented a system to recognize such patterns, at cellular and specimen levels, in images of HEp-2 cells. Ensembles of SVMs were trained to classify cells into six classes based on sparse encoding of texture features with cell pyramids, capturing spatial, multi-scale structure. Mean class accuracies for cell classification obtained on used test data sets were 87.1%. These were the highest achieved in the competition hosted by ICPR2014. At present, there are many mature image recognition and classification algorithms, and bag of features (BOF) algorithm is a popular one among them; BOF algorithm came from the bag of words (BOW) and was originally used for documents matching. In recent years, BOF algorithm is also been applied in the field of HEp-2 cell classification and has achieved results [12]. However, in the BOF algorithm there exists some problem in the application of image recognition and classification such as run is not fast enough and the classification accuracy is not high enough. So it needs to be optimized. Recent research has focused on extracting various morphological features and texture features such as local binary patterns (LBP) have been widely applied for classification of segmented images of HEp-2 cells [14]. Due to poor quality images, utilizing low-level features directly may be inefficient. Nosaka et al. [13] propose co-occurrence of adjacent local binary patterns (CoALBP) to extract textural features. Using linear support vector machine (SVM), their method won the first prize in the contest ICPR2012. Recently [15] a specimen-level image classification system was proposed, in which a specimen-level image descriptor is learned, based on cell-level attributes. In another system, proposed by Soda et al. [16], each specimen image is classified based on the strength of the recognition of the cells in that specimen image. Han et al. [17] have employed a parametric probability process to model local image patches (textons: microstructures in the cell image) and extract the higher-order statistics of the model parameters for the image description. They have used a simple linear support vector machine for cell pattern identification. Cascio et al. [18] proposed a classification approach based on two steps: the first step follows the one-against-all (OAA) scheme, while the second step follows the one-against- one (OAO) scheme. To do this, they have implemented 21 KNN classifiers: 6 OAA and 15 OAO. Perner et al. [19] presented an early attempt on developing an automated HEp-2 cell classification system. Cell regions were represented by a set of basic features extracted from binary images obtained at multiple grey level thresholds. Those features were then classified into six categories by a decision tree algorithm. Theodorakopoulos et al. [20] combined a set of local features, including LBP and rotation-invariant SIFT, with vectors of locally aggregated descriptors (VLAD).
Fig. 1. IIF images with different staining patterns (from left to right and from up to down: homogeneous, speckled, nucleolar, centromere, Golgi and nuclear membrane).
1.2. Our contributions In this paper we describe a system for the classification of pre-segmented immunofluorescence images of HEp-2 cells into six classes: homogeneous, speckled, nucleolar, centromere, Golgi, and nuclear membrane. Fig. 1 shows examples of each class. This system is based on multiple types of class-process which uses a twolevel pyramid to retain some spatial information. The basic intuition behind our approach is that, instead of using a single process to discriminate each class from all other classes, it is better to combine a set of different and complementary processes. For this purpose, an intensive analysis of preprocessing types has been conducted and for each process it has been identified the preprocessing method giving the best performance in terms of final class accuracy. That analysis was carried out by distinguishing between the mask cells and the adjacent circular crown, and 30 different best preprocessing were obtained. We extracted a large number of features (216), and for each classification process a phase of features reduction, based on Fisher criterion, was performed in order to select the best features, able to fully characterize the staining pattern of HEp-2 cells. We discuss the details of our approach in the following sections. As final consideration, we want to highlight that the method here presented has the advantage of allowing an easy extension to the problem of aided diagnosis, and hence to the development of a CAD for the support to diagnosis, which is the ultimate goal of this research [6], this is possible by including an automatic segmentation phase in the system, also differentiated (given the enormous variety of ways in which the pattern may occur, this is probably the only way to achieve an efficient system). 2. Materials and methods In several multiclass classification problems, it is preferable to use a number of classifiers equal to the number of classes and each classifier is trained in order to discriminate a class from all the others (binary approach) [21]. In this work, the classification stage is differentiated by using 15 one-against-one classifiers for six classes of staining patterns. In addition, the preprocessing and feature extraction steps are differentiated too. Therefore, we adopt a non-standard pipeline for supervised image classification. The generic image is simultaneously processed by 15 processes obtaining 15 separate outputs that represent how the cell resembles each one of the 6 classes analysed in this work. This system
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
JID: PATREC
ARTICLE IN PRESS
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
3
for each of the 30 OAO developed classifiers, obtained after an intensive and exhaustive analysis of the combinations of functions, taken two by two, reported in Table 1. The description of methods for selection of optimal preprocessing algorithms and the flow chart of the optimization method used for each (fifteen) classifiers one-against-one developed is shown in Section 2.4. As it can be easily seen from Table 2, the preprocessing algorithm obtained with contrast normalization and dilation has been associated with most types of OAO classifiers. 2.2. Features extraction and dimensionality reduction
Fig. 2. Pipeline of developed method.
uses a two-level pyramid to retain some spatial information. In fact, in addition to the segmented mask regions, the information is extracted from circular regions adjacent to the cells. For each image cell, the mask used for the segmentation was dilated (using a 7 × 7 structuring element) to obtain the adjacent circular crowncontaining information on the image area next to the cell. For each process a preprocessing phase is performed and a set of specific features are extracted both from these cells and from the adjacent circular crown. Finally a RBF support vector machine (SVM) classifier was employed to classify each image into one of two specific class patterns to which the specific classification process is addressed (binary approach, one-against-one). Fig. 2 shows the flow of operations adopted in this work. The following sections describe the proposed method in some more detail. The choice of methods, features and parameters was performed automatically, using the mean class accuracy (MCA) as a figure of merit. The parameters involved are tuned using a leave-onespecimen-out cross validation scheme. The main benefit of this pipeline is that it has a good explanatory power, as its principle is simple to explain.
The cells patterns analysis clearly shows that differences between classes are mainly based on the presence and distribution of bright/dark structures: numbers, sizes (areas), intensity and colours. It seems natural to use features providing an analysis of texture and more specifically features able to deal with bright/dark speckle-like structure description. Different staining patterns can be characterized by a limited set of attributes describing the spatial relationships between pixels values and the main image variations occurring in each cell type [22]; this information is generally obtained by mean of textural analysis techniques. These techniques can be grouped into two major categories: (i) statistical methods, describing the distribution of grey-levels in the image, and (ii) frequency domain measurements of image variations [23–25]. To achieve the objective of robust classification, we combined several discriminative visual features known to be effective for cell classifications with a robust and scalable multi-class boosting. All classifiers use a large number of extracted features, able to fully characterize the HEp-2 cells. In more details, four quantization levels [19, 26] were analysed: 8 bit, 7 bit, 6 bit and 5 bit of histogram intensity (respectively related to 256, 128, 64 and 32 Gy levels). For each quantization level the following 27 features have been extracted: •
•
•
2.1. Preprocessing The preprocessing phase in a pattern recognition task as that here described, where the type of objects to be classified is quite heterogeneous, is very important, and surely affects the next phase of features extraction, and hence the classification itself. This is the reason why, as above anticipated, in this work we decided to differentiate the preprocessing phase for each type of pattern to be recognized. For this purpose, starting from a set of functions, analyzed and reported in Table 1, many combinations of different types of preprocessing have been analysed and for each class it has been identified the preprocessing method giving the best performance in terms of final class accuracy. For many functions reported in Table 1, several different configurations have been evaluated, using several parameters (e.g. for the Gaussian filter the convolution mask, sized 3 × 3, 5 × 5 and 7 × 7 have been analysed), leading to a total of 37 analysed functions. Table 2 shows the details of combinations of preprocessing and their use in one-against-one classification processes. The analysis was conducted by distinguishing between the mask cells and the circular crown, for which, in fact, 30 different best preprocessing methods were obtained. In Table 2 the best results were reported
•
Intensity based features (9): mean value, standard deviation, ratio of the standard deviation to the mean value, entropy, moment of inertia, skewness, kurtosis, entropy of the contours gradient [27–29]. Geometry-based features (8): mean radius, standard deviation of radius, maximum radius, ratio of the standard deviation to the mean value, circularity, anisotropy, fractal index, eccentricity [30]. Shape-morphological-based features (8): area, perimeter, convex area, convex deficiency, solidity, compactness, roundness, Euler’s number. Descriptors-based features (2): entropy of HOG (histogram of oriented gradients), entropy of HAG (histogram of amplitude gradients).
From each cell image, all the features were densely extracted from image patches and from the circular crown obtained by mask dilation. To summarize, the 27 features above mentioned, being extracted using 4 quantization levels from image patches and from the circular crown, led to get a total analysis over 216 features (27 × 4×2 = 216). Reducing the dimensionality of the data by selecting a subset of the original variables may be advantageous to simplify the classification problem from different points of view. As the dimensionality of the data increases, many types of data analysis and classification problems become significantly harder. Sometimes the data also becomes increasingly sparse in the feature space. This can lead to big problems for both supervised and unsupervised learning. The main idea of feature subset selection is to remove redundant or irrelevant features from the dataset as they can negatively influence the classification accuracy, the clustering quality and lead to an unnecessary increase of computational cost. The
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC 4
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8 Table 1 List of functions analysed for preprocessing. Abbreviation
Function description
Reference
Nt Cn
Nothing Contrast normalization: was obtained by linearly remapping the intensity values so that 1% of data is saturated at low and high intensities. Equalization CLAHE (contrast limited adaptive histogram equalization) Median filter Gaussian filter Gaussian-enhancement: Dilatation Erosion Nagao Bilateral Opening FAS (filter alternate segmental)
http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://docs.opencv.org http://www.pkuwwt.tk/ofeli/doc/index.html http://docs.opencv.org http://docs.opencv.org http://www.pkuwwt.tk/ofeli/doc/index.html
Eq Ch Md Gs Ge Di Er Ng Bl Op Fs
Table 2 List of preprocessing used in OAO classifiers. Preprocessing for mask cell
Preprocessing for circular crown
N
Function 1
Function 2
Function 1
Function 2
Used for classification:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Cn Di Di Cn Cn Cn Cn Cn Cn Cn Cn Bl Cn Cn Op
Ng Cn Cn Bl Bl Bl Di Bl Bl Ge Ng Cn Bl Gs Cn
Di Di Di Eq Eq Di Cn Eq Cn Di Di Di Cn Di Cn
Cn Cn Cn Op Di Cn Fs Op Di Ch Cn Cn Gs Cn Gs
Homogeneous versus speckled Homogeneous versus nucleolar Homogeneous versus centromere Homogeneous versus Golgi Homogeneous versus nuclear membrane Speckled versus nucleolar Speckled versus centromere Speckled versus golgi Speckled versus nuclear membrane Nucleolar versus centromere Nucleolar versus Golgi Nucleolar versus nuclear membrane Centromere versus Golgi Centromere versus nuclear membrane Golgi versus nuclear membrane
advantage of identifying a subset of features is that no information about the importance of single features is lost. Filter methods perform feature selection in two steps. In the first step, the filter method assesses each feature individually for its potential in discriminating among classes in the data. In the second step, features falling beyond some threshold criterion are eliminated, and the smaller set of remaining features is used. This score-and-filter approach has been used in many recent publications, due to its relative simplicity [31,32]. Scoring methods generally focus on measuring the differences between distributions of features. The resulting score is intended to reject the quality of each feature in terms of its discriminative power. Many scoring criteria exist, in this work we have adopted the criterion of statistical selection of Fischer (Eq. (1)). By this criterion, the quality of the ith feature is expressed in terms of the difference among the empirical means of two distributions, normalized by the sum of their variances.
V (i ) =
(μ1 (i ) − μ2 (i ) )2 σ12 + σ22
(1)
In our work, μ1 (i) and μ2 (i) represent the average values of ith feature respectively on pattern 1 and on pattern 2 (OAO classification), while σ 1 and σ 2 represent the corresponding standard deviations. For each pair of patterns to be classified, the Fisher score is therefore evaluated on each extracted feature, and only the 15 highest values obtained from the cells and the 15 highest values obtained from their circular crowns (30 features out of 216) were selected. All the developed classifiers have 30 input, of course not necessarily the same.
2.3. Classification The classification is a crucial task for this problem and requires a training phase computationally expensive. The conventional way to deal with a multi-class problem is to decompose an M-class problem into a series of two-class problems. The two approaches commonly used are the One-Against-One (OAO) and One-AgainstAll (OAA) techniques. The OAA approach [33,34] represents the earliest and most commonly used multiclass approach and involves the division of an M class dataset into M two-class cases; OAA uses M classifiers where the ith classifier separates class i from all the remaining classes. In order to assign a class to a new sample, it is necessary to evaluate the output for all binary classifiers and, usually, choosing the pattern relative to the classifier that returned the highest classification value. The problem with the participation to the decision is assumed to be equally reliable, which is rarely the case. The OAO approach on the other hand involves constructing a classifier for each pair of classes resulting in M(M–1)/2 classifier; the OAO approach is more computationally intensive. When applied to a test point, each classification gives one vote to the winning class and the point is labeled with the class having most votes. One drawback of one-against-one method, however, arises when the results from the multiple classifiers are counted for the final decision without considering the competence of the classifiers. In this paper we propose an approach based on OAO approach. For each pair of class patterns one classifier is implemented (e.g. homogeneous versus speckled), for a total of 15 classifiers (see Fig. 2). The final decision is obtained counting the obtained votes by each class patterns and choosing the one with the
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
5
with the higher discriminating power, and this is done by selecting, according to the Fisher criterion, the best 15 features for the masks and the best 15 features for circular crowns. For each combination of preprocessing functions is evaluated both the value DM given by the sum of the 15 best values of Fisher for the masks, and the sum of the DC of the 15 best values of Fisher for the circular crowns.
DM =
15
Vmask (i )
(3)
Vcrown (i )
(4)
i=1
DC =
15 i=1
The selection of the best preprocessing functions for each process was carried out in order to maximize the sum DM + DC ; for each binary classification, the configuration having maximum DM is stored and at the end is coupled to the one with maximum DC . The configuration of functions and features thus obtained, for the masks and for the circular crowns, is used for the subsequent step of identification of the parameters of the classifier. In the end, for each binary classifier, it is identified a set of preprocessing functions, a set of features, and an SVM with the relevant parameters. The SVMs training was carried out to maximize the MCA, using the leave-one-specimen-out cross validation cross procedure. In the end the optimization method leads to a multi-process system as represented in the schema of Fig. 2. 2.5. Leave-one-specimen-out
Fig. 3. Flow chart of the optimization method used for each developed (fifteen) one-against-one classifier.
most votes. SVMs (Support Vector Machines) are a useful technique for data classification, because, unlike many classifiers in which it is easy to overfit the data, SVM provides a better generalization bound. In this work it was decided to use SVM classifiers. The (Gaussian) RBF (Radial Basic Function) kernel is a reasonable first choice. This kernel nonlinearly maps samples into a higher dimensional space, so that, unlike the linear kernel, can handle the case when the relation between class labels and attributes is nonlinear. Furthermore, the linear kernel is a special case of RBF [35,36] since the linear kernel with a penalty parameter C has the same performance as the RBF kernel with some parameters (C, γ ). In addition, the sigmoid kernel behaves like RBF for certain parameters [35,36]. In this work we adopted the following function kernel:
2
k(x, y ) = exp −γ xi − x j ,
γ >0
(2)
2.4. Method for functions and parameters selection In the present work an intensive analysis of methods and parameters was performed, in order to analyze and classify in different ways the patterns sought. This system indeed, for each binary classification, identifies: a set of preproccesing functions (4 functions), a set of features (30 features) and a set of parameters for the SVM classifier (C and γ ). In Fig. 3 it is shown the flow chart of the optimization method used for each of the 15 implemented processes. As it can be seen from Fig. 3, any two preprocessing functions between those implemented (shown in Table 1) are combined in a comprehensive manner and the related preprocessing is performed. The images thus obtained are used for the extraction of features, both for the masks and for the circular adjacent regions. The next phase of features has the aim of identifying the features
In this work the leave-one-specimen-out cell cross validation technique [11,37] have been used to exploit the highest possible number of patterns during the training phase, without invalidating the results. The method consists in leaving out all cells belonging to the same specimen, rather than leaving out a single cell for the construction of the training set; cells of the same specimen are similar in terms of the average intensity and contrast, and introduce bias. The leave-one-specimen-out strategy is an effective way to evaluate the algorithm when the test data set is not available. 2.6. Statistics calculation The Accuracy and the mean class accuracy are adopted in this work as measures of the performance [38]. Let CCRk be the correct classification rate for class k determined as follows:
CCRk =
Tk Nk
(5)
where Tk is the number of correct identifications of class k, while Nk is the total number of elements of class k. The accuracy is defined by:
Accuracy =
k
CCR · N k k k Nk
(6)
The MCA is determined by:
MCA =
1 CCRk k k
(7)
3. Results 3.1. Dataset The development of a CAD system is intimately linked to collection of a dataset of selected images [39]. For this study we used
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC 6
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8 Table 3 Best values for C and γ parameters.
Fig. 4. Distribution of training set patterns.
the dataset provided for the participation to the “contest on performance evaluation on indirect immunofluorescence image analysis systems”, hosted by the 22nd International Conference on Pattern Recognition (ICPR 2014); the dataset can be downloaded after proper registration on site http://i3a2014.unisa.it/. The participants are required to face exactly the same task of the ICIP 2013 Contest by using the same training and testing dataset [40]. The goal is the design and implementation of an IIF pattern recognition system able to classify the cells belonging to HEp-2 images in one of the following pattern classes: homogeneous, speckled, nucleolar, centromere, Golgi and nuclear membrane. The dataset has been collected between 2011 and 2013 at Sullivan Nicolaides Pathology laboratory, Australia. It utilizes 419 patient positive sera, which were prepared on the 18-well slide of HEP-20 0 0 IIF assay from Immuno Concepts N.A. Ltd. with screening 1:80 dilution. The specimens were then automatically photographed using a monochrome high dynamic range cooled microscopy camera which was fitted on a microscope with a plan-Apochromat 20x/0.8 objective lens and an LED illumination source. Approximately 100– 200 cell images were extracted from each patient serum. In total there were 13,596 cell images used for training. The labelling process involved at least two scientists who read each patient specimen under a microscope. A third expert’s opinion was sought to adjudicate any discrepancy between the two opinions. We used each specimen label for the ground-truth of cells extracted from it. Furthermore, all the labels were validated by using secondary tests such as ENA and anti-ds-DNA in order to confirm the presence and/absence of specific patterns. Each cell image contained in the database is annotated with the following information: cell pattern (one of the patterns above defined), cell intensity, cell mask, ID of the image which the cell belongs to. A characterization of the dataset is reported in Fig. 4. 3.2. SVM parameters There are two parameters for an RBF kernel: C and γ . It is not known beforehand which C and γ are best for a given problem; consequently some kind of model selection (parameter search) must be done. The goal is to identify good (C; γ ) so that the classifier can accurately predict unknown data (i.e. testing data). Note that it may not be useful to achieve high training accuracy (i.e. a classier which accurately predicts training data whose class labels are indeed known). A common strategy is to separate the dataset into two parts, of which one is considered unknown. The prediction accuracy obtained from the “unknown" set more precisely reflects the performance on classifying an independent dataset. We have used a “grid-search" on C and γ using leave-one-specimenout validation. Various pairs of (C, γ ) values are tried and the one
Class
C
γ
Homogeneous versus speckled Homogeneous versus nucleolar Homogeneous versus centromere Homogeneous versus Golgi Homogeneous versus nuclear membrane Speckled versus nucleolar Speckled versus centromere Speckled versus Golgi Speckled versus nuclear membrane Nucleolar versus centromere Nucleolar versus Golgi Nucleolar versus nuclear membrane Centromere versus Golgi centromere versus nuclear membrane Golgi versus nuclear membrane
8 8 1024 4 32 4 4 4 4 4 4 64 4 32 4
1.0 0.5 7.8 × 10−3 7.8 × 10−3 1.3 × 10−1 7.8 × 10−3 3.1 × 10−2 1.6 × 10−2 1.3 × 10−1 7.8 × 10−3 7.8 × 10−3 1.6 × 10−2 1.0 1.6 × 10−2 3.1 × 10−2
Table 4 Per-class accuracy. Class
Accuracy
Centromere Golgi Homogeneous Nucleolar Nuclear membrane Speckled
87.09% 88.54% 93.70% 89.26% 90.24% 88.70%
with the best accuracy is taken. We found that trying exponentially growing sequences of C and γ is a practical method to identify good parameters [35,36]. In this paper, the probed values for C and γ were:
C = 2−5 , 2−4 , . . . , 210 ; γ = 2−10 , 2−9 , . . . , 22 For each pair of class pattern the analyzed grid had size equal to 16 × 13, i.e. 208 points. In Table 3 are reports the best values of parameters C and γ obtained for each pair of class pattern. 3.3. Experimental results The proposed approach has been trained on the dataset provided by the ICPR 2014 contest for task 1. Task 1 is delegated to cell level classification. In particular, it consists of 13,596 images for the training phase coming from 419 patient positive sera and categorized by trained physicians into six patterns: centromere, homogeneous, nucleolar, speckled (coarse and fine), nuclear membrane and Golgi. The classification results obtained in our experiments have been summarized in Table 4, and organized by staining pattern class; where, for each of them, we show the obtained accuracy. The estimated mean class accuracy for the proposed method is 90.25%, with a maximum per-class accuracy of 93.7% for homogeneous patterns. From the table we can notice that the poorest classification accuracy is achieved by centromere pattern class. For some patterns, such as the Golgi, the improvement in classification performance was significant, due to use of the information contained in circular crowns. Furthermore, the software implementation of the method presented here has been submitted to the group owning the database used in ICPR2014 contest [41], and this allowed to evaluate the system’s performance on the same “test set” used in task 1 of ICPR2014 contest; this dataset was not used in training phase. The performances results are reported in the confusion matrix presented in the Table 5. The mean class accuracy (MCA) obtained is equal to 80.12%, with a maximum per-class accuracy of 93.7% for homogeneous
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
ARTICLE IN PRESS
JID: PATREC
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8 Table 5 Confusion matrix of our method.
Centr Golgi Homog Nucle NucMe Speckl
Centr
Golgi
Homog
Nucle
NucMe
Speckl
93.7% 0.03% 0.02% 1.6% 0.2% 15.0%
0,2% 76.4% 0.7% 0.7% 2.3% 0.5%
0.7% 2.0% 73.8% 2.2% 5.2% 13.3%
2.3% 8.1% 4.4% 85.1% 4.3% 2.2%
0.3% 11.3% 13.1% 2.6% 86.7% 3.9%
2.9% 2.2% 8.0% 7.8% 1.5% 65.2%
Fig. 5. Mean accuracy obtained by all the methods on the test set ICPR2014.
patterns, and a minimum per-class accuracy of 65.2% for speckled patterns. In order to have a comparison with other developed methods presented in other works, a graph (Fig. 5) showing the performance obtained, on the same dataset, by other research groups at the ICPR 2014 context, is following reported. As it can be easily deduced by Fig. 5, the method presented in this work is definitely comparable, in terms of performance, with the results obtained by other research groups in the recent works. 3.4. Running time The average running time to analyse a cell image and its mask is 0.8 s, about 2 min per IIF image (using a 3.4 GHz Intel i7 CPU). The training time, using Fisher analysis, for each classifier, takes about 4 h. The tuning of SVMs takes about 45 min for each classifier. Note that it is possible to reduce the total running time by parallelizing each of the fifteen independent processes. 4. Discussion and conclusions In this paper we proposed an approach for the automatic classification of staining patterns in HEp-2 cell IIF images, which is critical for the diagnosis of autoimmune diseases. We adopted a non-standard pipeline for supervised image classification that provides a diversified process for all types of class-patterns to classify. For this purpose, some different types of preprocessing algorithms have been analysed and for each process it has been identified the preprocessing method giving the best performance in terms of final class accuracy. We extracted a large number of features (216), able to fully characterize the staining pattern of HEp-2 cells. Then, for each class, a sub-set of 30 features has been selected. Furthermore, we proposed a classification-approach based on OAO classification rule. Our classification system is thus composed by 15 SVM classifiers, with an approach one-against-one. Leave-one-specimenout cross validation method was used for the optimization of the system. We evaluated the method performances on the HEp-2 cells dataset (manually segmented and annotated). Although being the
early results of our methodology for such a challenging application, performances are really satisfactory (mean class accuracy of 80.12%). The choice to combine the OAO approach, therefore, produces an improvement, even if not notable, of the mean class accuracy. The significant difference between the results, presented in the Table 3, obtained with leave-one-out method and those of the confusion matrix (Table 4) is certainly to ascribe to the presence of bias. The present method has certainly the advantage that can be easily extended to research for additional autoantibody patterns having simply to add, in this case, the classification processes related to possible new patterns to search. In addition, the intensive analysis conducted regarding preprocessing functions and parameters can be used in order to get information for a study based on LDA or similar methods. The used method of reduction of dimensionality, compared to other criteria, highly reduces the computational load and, in that sense, allowed to perform the intensive analysis of functions and parameters carried out for each of the 15 implemented classification processes. The major limitation of the method is certainly linked to the lack of correlation analysis between the features (that, on the contrary, the LDA is able to do); two totally related features may be both selected by the Fisher method. The performance of the proposed method can be easily improved. In fact, the method of coupling the results of the 15 classifiers obtained as simple count of votes, is a simplistic solution, probably the inclusion of a final classifier (such as K-NN) could have better enhanced the conducted analysis. Further analysis is being performed to evaluate the robustness and reliability of other features selection methods. This includes the evaluation of exhaustive or almost-exhaustive selection approaches, by focusing on the feature space proposed in this work.
References [1] National Institutes of Health, The autoimmune disease coordinating committee, Progress Autoimmune Dis. Res. (2005) 1–146. [2] X. Xiang, L. Feng, C. Ng, L.K. Pang, Staining pattern classification of ANA-IIF based on SIFT features, J. Med. Imag. Health Inf. 2 (4) (2012) 419–424. [3] S. Ciatto, et al., Computer assisted diagnosis (CAD) in mammography. Comparison of diagnostic accuracy of a new algorithm (Cyclopus®, Medicad) with two commercial systems, Radiol. Med. 114 (4) (2009) 626–635. [4] D. Cascio, F. Fauci, M. Iacomi, G. Raso, R. Magro, D. Castrogiovanni, G. Filosto, R. Ienzi, M. Vasile, Computer-aided diagnosis in digital mammography: comparison of two commercial systems, Imag. Med. 6 (1) (2014) 13–30. [5] A. Rigon, F. Buzzulini, P. Soda, L. Onofri, L. Arcarese, G. Iannello, A. Afeltra, Novel opportunities in automated classification of antinuclear antibodies on HEp-2 cells, Autoimmunity Rev. 10 (10) (2011) 647–652. [6] A. Elgaaied Benammar, et al., Computer-assisted classification patterns in autoimmune diagnostics: the A.I.D.A. Project, Biomed. Res. Int. 2016 (2016) 1–9. [7] P. Foggia, G. Percannella, P. Soda, M. Vento, Benchmarking hep-2 cells classification methods, IEEE Trans. Med. Imag. 32 (10) (2013) 1878–1889. [8] P. Foggia, G. Percannella, A. Saggese, M. Vento, Pattern recognition in stained HEp-2 cells: Where are we now? Pattern Recognit. 47 (7) (2014) 2305–2314. [9] [9] Lovell, B.C., Percannella, G., Vento, M., Wiliem, A., 2014, Performance Evaluation of Indirect Immunofluorescence Image Analysis Systems, Technical Report, ICPR Workshop. [10] S. Ensafi, S. Lu, A.A. Kassim, C.L. Tan, in: A bag of words based approach for classification of hep-2 cell images. Pattern recognition techniques for indirect immunofluorescence images (I3A), 2014 1st Workshop, 2014, pp. 29–32. [11] S. Manivannan, W. Li, S. Akbar, R. Wang, J. Zhang, S. McKenna, An automated pattern recognition system for classifying indirect immunofluorescence images of HEp-2 cells and specimens, Pattern Recognit. 51 (2016) 12–26. [12] L. Shen, J. Lin, S. Wu, S. Yu, HEp-2 image classification using intensity order pooling based features and bag of words, Pattern Recognit. 47 (7) (2014) 2419–2427. [13] R. Nosaka, K. Fukui, HEp-2 cell classification using rotation invariant co-occurrence among local binary patterns, Pattern Recognit. 47 (7) (2014) 2428–2436. [14] A.B.L Larsen, J.S. Vestergaard, R. Larsen, HEp-2 cell classification using shape index histograms with donut-shaped spatial pooling, IEEE Trans. Med. Imag. 33 (7) (2014) 1573–1580. [15] B.C.L.A. Wiliem, P. Hobson, Discovering discriminative cell attributes for hep-2 specimen image classification, in: IEEE Winter Conference on Applications of Computer Vision, Article number 6836071, 2014, pp. 423–430.
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024
7
JID: PATREC 8
ARTICLE IN PRESS
[m5G;April 25, 2016;22:24]
D. Cascio et al. / Pattern Recognition Letters 000 (2016) 1–8
[16] P. Soda, G. Iannello, Aggregation of classifiers for staining pattern recognition in antinuclear autoantibodies analysis, IEEE Trans. Inf. Technol. Biomed. 13 (3) (2009) 322–329. [17] X.-H. Han, J. Wang, G. Xu, Y.-W. Chen, High-order statistics of microtexton for hep-2 staining pattern classification, IEEE Trans. Biomed. Eng. 61 (2014) 2223–2234. [18] D. Cascio, V. Taormina, M. Cipolla, F. Fauci, M. Vasile, G. Raso, HEp-2 cell classification with heterogeneous classes-processes based on K-nearest neighbours, in: IEEE Computer Society – Proceedings of 1st Workshop on Pattern Recognition Techniques for Indirect Immunofluorescence Images, 2014, pp. 10–15. [19] P. Perner, H. Perner, B. Muller, Mining knowledge for Hep-2 cell image classification, Artif. Intell. Med. 26 (1–2) (2002) 161–173. [20] I. Theodorakopoulos, D. Kastaniotis, G. Economou, S. Fotopoulos, HEp-2 cells classification using morphological features and a bundle of local gradient descriptors, in: 1st Workshop on Pattern Recognition Techniques for Indirect Immunofluorescence Images (I3A), 2014, pp. 33–36. [21] J Furnkranz, Round Robin classification, J. Mach. Learn. Res. 2 (4) (2002) 721–747. [22] A. Sriram, S. Ensafi, S. Roohi, A.A. Kassim, Classification of human epithelial type-2 cells using hierarchical segregation, in: 13th International Conference on Control Automation Robotics Vision (ICARCV), 2014, 2014, pp. 323–328. [23] G.L. Masala, et al., Comparative study of feature classification methods for mass lesion recognition in digitized mammograms, Nuovo Cimento C 30 (3) (2007) 305–316. [24] L. Vivona, et al., Fuzzy technique for microcalcifications clustering in digital mammograms, BMC Med. Imag. 14 (1) (2014) 1–14. [25] G.L. Masala, B. Golosio, et al., Classifiers trained on dissimilarity representation of medical pattern: A comparative study, Nuovo Cimento C 28 (6) (2005) 905–912. [26] M. Iacomi, et al., Mammographic images segmentation based on chaotic map clustering algorithm, BMC Med. Imag. 14 (1) (2014) 1–11. [27] L. Vivona, et al., A fuzzy logic C-means clustering algorithm to enhance microcalcifications clusters in digital mammograms, in: Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) IEEE, Article number 6152551, 2011, pp. 3048–3050.
[28] F. Fauci, et al., A Fourier based algorithm for microcalcifications enhancement in mammographic images, in: IEEE Nuclear Science Symposium and Medical Imaging Conference, Article number 4774254, 2012, pp. 4388–4391. [29] D. Cascio, et al., Mammogram segmentation by contour searching and mass lesions classification with neural network, IEEE Trans. Nucl. Sci. 53 (5) (2006) 2827–2833. [30] D. Cascio, R. Magro, F. Fauci, M. Iacomi, G. Raso, Automatic detection of lung nodules in CT datasets based on stable 3D mass-spring models, Comput. Biol. Med. 42 (11) (2012) 1098–1109. [31] P. Duda, D.G. Stork, Pattern Classification, Wiley-Interscience Publication, 2001. [32] Q. Gu, Z. Li, J. Han, Generalized Fisher score for feature selection, in: Proceedings of the 27th Conference on Uncertainty in Artificial Intelligence, 2011, pp. 266–273. [33] B. Liu, Z. Hao, E.C.C. Tsang, Nesting one-against-one algorithm based on SVMs for pattern classification, IEEE Trans. Neural Netw. 19 (12) (2008) 2044–2052. [34] R. Debnath, N. Takahide, H. Takahashi, A decision based one-against-one method for multi-class support vector machine, Pattern Anal. Appl. 7 (2) (2004) 164–175. [35] C.-C. Chang, C.-J. Lin, LIBSVM: a Library for Support Vector Machines, ACM Trans. Intell. Syst. Technol. 2 (3) (2011) 1–27, Article 27. [36] C.-W. Hsu, C.-C. Chang, C.-J. Lin, A Practical Guide to Support Vector Classification, Department of Computer Science National Taiwan University, Taipei, Taiwan, 2003. [37] A.B.L. Larsen, J.S. Vestergaard, R. Larsen, HEp-2 cell classification using shape index histograms with donut-shaped spatial pooling, IEEE Trans. Med. Imag. 33 (7) (2014) 1573–1580. [38] A. Rigon, F. Buzzulini, P. Soda, G. Iannello, Afeltra, A Novel opportunities in automated classification of antinuclear antibodies on HEp-2 cells, Autoimmunity Rev. 10 (10) (2014) 647–652. [39] S. Tangaro, et al., MAGIC-5: an Italian mammographic database of digitized images for research, Radiol. Med. 113 (4) (2008) 477–485. [40] P. Hobson, B. Lovell, G. Percannella, M. Vento, A. Wiliem, Benchmarking human epithelial type 2 interphase cells classification methods on a very large dataset, Artif. Intell. Med. 65 (3) (2015) 239–250. [41] Techniques for Indirect Immunofluorescence Images (I3A), 2014 ICPR2014 contest http://nerone.diem.unisa.it/i3a/wp-content/uploads/ICPR2014_Report.pdf
Please cite this article as: D. Cascio et al., A multi-process system for HEp-2 cells classification based on SVM, Pattern Recognition Letters (2016), http://dx.doi.org/10.1016/j.patrec.2016.03.024