- Email: [email protected]
An adaptive positivity thresholding method for automated Ki67 hotspot detection (AKHoD) in breast cancer biopsies
Accepted Manuscript Title: An adaptive positivity thresholding method for automated Ki67 hotspot detection (AKHoD) in breast cancer biopsies Author: David Pilutti Vincenzo Della Mea Enrico Pegolo Francesco La Marra Fulvio Antoniazzi Carla Di Loreto PII: DOI: Reference:
S0895-6111(17)30036-8 http://dx.doi.org/doi:10.1016/j.compmedimag.2017.04.005 CMIG 1506
To appear in:
Computerized Medical Imaging and Graphics
Received date: Revised date: Accepted date:
19-7-2016 18-4-2017 23-4-2017
Please cite this article as: David Pilutti, Vincenzo Della Mea, Enrico Pegolo, Francesco La Marra, Fulvio Antoniazzi, Carla Di Loreto, An adaptive positivity thresholding method for automated Ki67 hotspot detection (AKHoD) in breast cancer biopsies, (2017), http://dx.doi.org/10.1016/j.compmedimag.2017.04.005 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ip t
An adaptive positivity thresholding method for automated Ki67 hotspot detection (AKHoD) in breast cancer biopsies
cr
David Piluttia , Vincenzo Della Meaa , Enrico Pegolob , Francesco La Marrac , Fulvio Antoniazzic , Carla Di Loretoc a
an
us
Department of Mathematics, Computer Science, and Physics, University of Udine, Udine, Italy b Institute of Pathology, University Hospital ”Santa Maria della Misericordia”, Udine, Italy c Department of Medical and Biological Science, University of Udine, Udine, Italy
M
Abstract
Ac ce p
te
d
The proliferative activity of breast cancer tissue can be estimated using the Ki67 biomarker. The percentage of positivity of such biomarker is correlated with proliferation and consequently with the prognosis of a breast tumor. Ki67 marked tissue samples are analyzed by an experienced pathologist who identifies the most active areas of tumor cell proliferation called hotspots, and estimates the positivity of each case. A method for the Automated Ki67 Hotspot Detection (AKHoD) is presented in this work. The main objective of the AKHoD method is to automatically and efficiently provide the pathologist with suggestions about Ki67 hotspot areas as a decision support. The input of AKHoD is a digital slide that is divided in tiles. For each tile, AKHoD provides a rough estimate of positivity and cellularity, summarized in very low resolution positivity and cellularity images. In a second step, an adaptive thresholding is applied to such positivity image to identify the most positive connected and convex areas, within cellularity limits set by current guidelines (that is, 500-2000). The method has been preliminarily validated on 50 digital slides for which three expert pathologists provided gold standard hotspots. 82% of the gold standard hotspots have been successfully recognized by the system, spending an average of 54 seconds per slide. While further validation is needed taking into account also patients followup, this first experimentation suggests that the proposed method could be adequate for supporting the pathologist in Preprint submitted to Computerized Medical Imaging and Graphics
April 18, 2017
Page 1 of 17
hotspot detection.
ip t
Keywords: Digital Pathology, Ki67, Automated Hotspot Detection, Breast Cancer, Image Analysis
cr
1. Background
Ac ce p
te
d
M
an
us
The identification of hotspots as well as the quantification of immunohistochemical Ki67 positive staining are of critical importance for the prognosis and the treatment planning of breast cancer. A high percentage of Ki67 positivity within a biopsy indicates an higher proliferation rate and consequently the aggressiveness of a breast tumor. The proliferation rate is normally visually estimated by the pathologist over an hotspot area that should include from 500 to 2000 cells [1, 2]. This makes the identification and quantification of the hotspot areas as well as the calculation of the proliferation rate complicated and time consuming. To address such tasks, different approaches have been used. A semiautomated method for the identification of Ki67 hotspots has been proposed [3]. This method reduces the zooming operations that a pathologist must perform to select a hotspot area by enlarging positive nuclei at low magnification to make them easier to identify. ImmunoRatio is an open source web based application which performs a calculation of positively stained nuclear area by applying color deconvolution and adaptive thresholding [4]. The input is represented by stained images acquired with digital cameras and compressed in common formats such as JPEG, JPEG2000, TIFF, BMP, and PNG. The execution of ImmunoRatio generates a pseudo-colored segmented image which can be successively analyzed by an expert. Fully automated methods for the identification of Ki67 hotspot areas have also been developed. A method for the detection of Ki67 hotspot areas is the Automated Selection of Hotspots (ASH) [5], a Linux desktop application. The input of the ASH method is a Hamamatsu NDPI image which is splitted into smaller images called image blocks. Each image block is converted in a common image format such as TIFF or JPEG and analyzed using ImmunoRatio [4] to rank the top 10 image blocks, which represent potential hotspots. Clustering methodologies have also been used to address the hotspot detection in Ki67 stained neuroendocrine tumor images [6]. In such clustering framework the positive cell nuclear sections are segmented using minimal 2
Page 2 of 17
te
d
M
an
us
cr
ip t
graph cuts. The resulting segmentation is then used to detect the hotspots using non linear transformation for the clustering, allowing inclusions of positive nuclear sections in a cluster even in the case of sparse nuclei. A learning algorithm is combined with the clustering framework to simulate the human perception of an hotspot. An alternative method [7] detects Ki67 hotspots using hexagonal tiling and implementing the concept of Pareto hotspot, which represents the upper quintile of the biomarker expression in the tissue, providing also a quantitative analysis over the detected hotspots. The hotspot detection for Ki67 stained slides of breast cancer has also been posed as an image filtering problem [8]. This approach considers all the pixels corresponding to positive stained nuclei as a surrogate representation of a nuclear section. A lowpass filter is then applied to such image to obtain an approximated density map where the local maxima contribute to the identification of the hotspots in the image. In this paper a new efficient and robust method for the Automated Ki67 Hotspot Detection (AKHoD) is proposed with the objective to provide the pathologists with a fast and robust method for decision support applied to the analysis of breast cancer biopsies. The main goal of the AKHoD method is to automatically identify hotspot areas over Ki67 stained breast cancer biopsies images and provide the pathologist with visual suggestions. An approximated estimation of the proliferation rate as well as an estimation of cell nuclear section is also calculated within the suggested hotspot areas.
Ac ce p
2. Materials and Methods
The digital slides used for this study have been previously acquired as material for the AIDPATH breast cancer database [9]. On 50 invasive breast cancer biopsy samples immunohistochemistry was performed using the Ki67/MIB1 antibody, stained with Diaminobenzidine (DAB) and counterstained with Hematoxylin (H), resulting in H/DAB staining. The samples have been acquired using an Aperio CS (Leica Biosystems, Nussloch, Germany) slide scanner at 40x, with a resolution of 0.2491µm per pixel. Slides were saved in TIFF using the proprietary Leica SVS format. Input Whole Slide Images (WSI) are divided in tiles, whose colors are deconvolved using the method of Ruifrok [10] using the standard Hemathoxylin/Diaminobenzidine (H/DAB) deconvolution matrix and assuming a certain constancy in sample staining and preparation. This results into two
3
Page 3 of 17
Ac ce p
te
d
M
an
us
cr
ip t
deconvolved images which express the H and the DAB color component, respectively. The presence of Ki67 is quantified by calculating a ratio between the DAB and the H component. The number of tiles varies depending on the magnification level selected for the analysis. This results in a grouping of tiles that reduces the computation time needed for the analysis. The result is a positivity image in which each pixel corresponds to a tile in the input image and the intensities are the calculated DAB over H average positivity ratios. A dynamic color threshold is applied to such image to highlight potential hotspot areas. The optimization procedure starts from the maximum normalized density value used as color threshold that is iteratively diminished until the found hotspot areas have a sufficient size to contain an adequate number of nuclear sections. The stop criterion for iterations has been set to 2000 nuclear sections. The resulting hotspot areas are then written in an OME-XML [11] file which can be read by a medical image viewer such as Aperio ImageScope. To provide the pathologist with more significative suggestions, the three hotspot areas with highest DAB/H ratio are highlighted using a different color. The DAB/H ratio is calculated within each hotspot area and the result is also provided in the output. Hotspot areas have been identified and annotated over the 50 considered cases by three experienced pathologists. The manual annotations have been compared with the results from the AKHoD method. A percent overlap between hotspot areas from the experts and from AKHoD has been calculated. This gives a quantitative measure of the quality of the results obtained with the AKHoD method.
2.1. Positivity Quantification and Cellularity Estimation The pixels Hp and DABp corresponding to the H and DAB channel, respectively, are counted and the ratio (DABp /Hp )∗100 is calculated, resulting in a percent estimation of the DAB staining within a tile. This contributes to create a positivity image IP OS in which each pixel corresponds to a tile and the intensities correspond to the estimated DAB densities. In Fig. 1 an example of a WSI image stained with H/DAB and the corresponding positivity image obtained is shown, where a high intensity corresponds to a high DAB staining. The quantification of the DAB staining has to be integrated with an estimation of the number of cell nuclear sections within a tile to provide a more reliable result for determining the size of an hotspot. The average 4
Page 4 of 17
te
d
M
an
us
cr
ip t
radius for a breast tumor nuclear section is assumed to be 10µm. This value is used to statistically estimate the number of DAB stained nuclear sections by dividing the area corresponding to the DAB staining with the average nuclear area. This results in a cellularity image INDAB where the intensity of each pixel represents the number of the counted nuclear sections. A similar estimation is performed for the H staining assuming an average size of the cell nuclear section of 7µm, with the exclusion of double colored areas to obtain the image INH . In both estimations the background area is also excluded. In Fig. 2 there are examples of the nuclei density estimation of DAB and H staining components.
(b) positivity image
Ac ce p
(a) WSI image
Figure 1: Example of a WSI image of a breast tumor biopsy stained with H/DAB and its corresponding positivity image. In (b) each pixel correspond to a tile in (a). The intensities in (b) represent the DAB/H density, where an high intensity corresponds to a high DAB staining.
2.2. Hotspots Estimation The estimation of the hotspots is performed over IP OS with the support of INDAB and INH . The positivity image IP OS is thresholded starting from the threshold td = 100 which is the maximum value possible. The thresholded image is used to segment positive regions, calculate the relative areas with the average positivity percent value. This results in a table in which each line represents a different segmented region of the positivity image, providing 5
Page 5 of 17
ip t cr us (b) H nuclei
an
(a) DAB nuclei
M
Figure 2: Example of cellularity images. In (a) is an estimated DAB stained cellularity image. In (b) is an estimated H stained cellularity image with the exclusion of the double stained pixels. The intensity represents the number of the counted nuclear sections.
Ac ce p
te
d
also information about the position and the size of each region. The position and the size of the resulting regions are then also localized in INDAB and INH to perform a nuclear section count within the selected area. The regions containing a total number of nuclear sections nc below 500 are excluded from the results table. The AKHoD method iterates by lowering td and consequently progressively analyzing bigger parts of IP OS until one candidate hotspot area contains more than 2000 nuclear sections or reaches the maximum number of iterations k = 100. In Fig. 3 there is an example of a positivity image IP OS of Fig. 1b to which three different thresholds td are applied, highlighting different areas of IP OS . At the end of the iterations, the resulting candidate hotspot areas are classified with respect of their average DAB positivity value and written in the OME-XML output file. In Fig. 4 there is an overview of the AKHoD method with the separation between the preparatory phase of defining ID , INDAB , and INH and the iterations with different td. If AKHoD identifies multiple candidate hotspot areas, the three areas of higher DAB positivity expression are highlighted using a red color. Other candidates are highlighted using the green color to differentiate them from the areas of higher DAB positivity expression.
6
Page 6 of 17
ip t (b) positivity td = 90
image,
cr
image,
(c) positivity td = 80
image,
us
(a) positivity td = 100
Ac ce p
te
d
M
an
Figure 3: An example of a positivity image to which are applied different thresholds td resulting in the highlight of progressively bigger areas of the image when lowering td.
Figure 4: Overview of AKHoD method. The positivity image ID is calculated in the preparatory phase together with INDAB and INH . The iterations are performed decreasing the value of td until the match between the estimations of DAB and the number of cell nuclear sections satisfy the definition of an hotspot.
7
Page 7 of 17
Ac ce p
te
d
M
an
us
cr
ip t
2.3. Validation Three experienced pathologists reviewed the Ki67 stained glass slides that were previously digitally acquired and marked the hotspot areas. These marked areas were then reported on the digital slides as annotations using the Aperio ImageScope software. Annotated digital slides provided a gold standard for evaluation. The validation has been performed semi-quantitatively by calculating the superimposition between the suggested candidate hotspots and the gold standard. A gold standard hotspot has been considered recognized by the system if, at visual analysis, there is significant superimposition between both. In fact, since both gold standard and calculated hotspots are rough shapes, having full correspondence is almost impossible. A gold standard hotspot has been considered ”missed” if there is no superimposition with calculated hotspots. A calculated hotspot is considered ”added” if there is no corresponding gold standard hotspot. The number of recognized, missed, and added hotspots is calculated for each slide. Since one aim of the present work is to provide a fast system for decision support to the pathologist, the AKHoD method has been tested on different series (40x, 10x, 2.5x) included in the slide file, using also different tile sizes (60, 120, 240). The best compromise between recognition capabilities and speed has been then chosen. After the selection of one set of parameters, tests have been performed on 3 different machines, recording the time needed for execution on each slide to evaluate the average time performance. Execution time has been compared with image and file size using Pearson correlation.
2.4. Implementation The AKHoD method has been implemented in Java using the BioFormats library [12] to read WSI images. The tiles have been deconvolved using the color deconvolution method of Ruifrok [10] implemented in ImageJ. The functions of thresholding and segmentation of positive regions with ParticleAnalyzer have been integrated into the main code by using ImageJ as Java library [13, 14, 15]. The output XML files containing the hotspot suggestions have been created using the OME-XML structure [11]. The testing has been performed on three different machines, two workstations and a server.
8
Page 8 of 17
3. Results
Ac ce p
te
d
M
an
us
cr
ip t
The patient datasets used in this work have been previously acquired in the context of clinical studies which were approved by the institutional ethical committee, and informed consent was obtained from all patients. All the datasets have been also anonymized before testing. All the datasets have been provided with an indication of gold standard hotspots made by experienced pathologists. In Fig. 5 there are the comparative results of gold standard and AKHoD results in a case with positivity of 30% resulting in a good match between gold standard hotspots and AKHoD suggestions.
(a) Gold standard
(b) AKHoD suggestions
Figure 5: Comparative examples between gold standard hotspots (a) and the suggested areas from AKHoD (b) over a case of etimated positivity of 30%. All the hotspots area from the gold standard have been detected also from the AKHoD method.
In Fig. 6, instead, the AKHoD method has been tested on a case with positivity of 7%. It can be seen that both gold standard hotspot have been 9
Page 9 of 17
te
d
M
an
us
cr
ip t
also detected from the AKHoD method, which adds another area that have to be further validated by the pathologist.
(b) AKHoD suggestions
Ac ce p
(a) Gold standard
Figure 6: Comparative examples between gold standard hotspots (a) and the suggested areas from AKHoD (b) over a case of estimated positivity of 7%. Both of the hotspots areas from the gold standard have been detected also from the AKHoD method. A third hotspot area has been detected only from the AKHoD method, providing information that has to be further validated from the expert pathologist.
Finally, in Fig. 7 (65% positivity) one of the suggestions from the AKHoD method matched a gold standard hotspot, while the second suggestion resulted in a merge of two adjacent gold standard hotspot areas. The preliminary optimization of the algorithm parameters allowed to choose the series 2 (2.5x) with a tile size of 60 pixels as the best compromise between recognition performance and speed. In fact, the best one was series 0 (40x) with tile size 240 pixels, that recognized one hotspot more than the selected one, but with an average time needed for recognition 15 times higher. 10
Page 10 of 17
ip t cr us an M
(a) Gold standard
(b) AKHoD suggestions
te
d
Figure 7: Comparative examples between gold standard hotspots (a) and the suggested areas from AKHoD (b) over a case of estimated positivity of 65%. In this case one hotspot area has been correctly detected from the AKHoD method while the other is the result of the merge between two adjacent hotspot areas in the gold standard image.
Ac ce p
The AKHoD method has been run on the above mentioned slides, correctly recognizing 59 out of 72 gold standard hotspots (81.94%) on which all the three experts agree. Taking into consideration the hotspots on which at least two pathologists agree, the AKHoD system correctly recognized 84 hotspots out of 107 (78.50%). Finally, considering the hotspots on which at least one pathologist agree, the AKHoD system correctly recognized 118 hotspots out of 172 (60.61%). In some cases, system-calculated hotspots were overlapping with two adjacent gold standard hotspots. Table 1 shows details about correctly recognized hotspots and compare them with the agreement of the pathologists. Some statistical indexes such as Dice, Jaccard, sensitivity, and precision have also been calculated by comparing the number of hotspots recognized by the system and the hotspot areas provided by the pathologists. The overlapping hotspots have been considered as the True Positive (TP) values, while the False Positive (FP) are the hotspots recognized by the AKHoD
11
Page 11 of 17
Pathologists hotspots % 72 81.94 107 78.50 172 60.61
ip t
Agreement AKHoD n. Pathologists correct hotspots 3 59 2 84 1 118
us
cr
Table 1: Table presenting the number of the correctly recognized hotspots from the AKHoD system in comparison with the hotspots identified by the pathologists over the whole dataset. The rows provide the number of hotspots recognized by the AKHoD system over the number of hotspots on which at least three, two, and one patologists agree, respectively.
Pathologist 2 0.764 0.618 0.752 0.776
te
d
Index Pathologist 1 Dice 0.770 Jaccard 0.628 Sensitivity 0.784 Precision 0.757
M
an
system but not from the pathologists, and finally the False Negative (FN) are the hotspots that have been recognized by the pathologists but not by the system. As shown in Table 2, the average Dice coefficient resulted in 0.751 and the Jaccard in 0.602 with a sensitivity of 0.747 and a precision of 0.755. Pathologist 3 0.720 0.562 0.706 0.733
Average 0.751 0.602 0.747 0.755
Ac ce p
Table 2: Table describing the performance of the AKHoD method in comparison with the results obtained from the experts.
Although the prototype implementation was not yet optimized, the AKHoD method has been tested in 3 different systems to preliminary evaluate its average time performance: • MacBook 13, 2-core 2,8 GHz Intel Core i7, 16GB RAM, 512MB SSD running Windows 7: 84 seconds; • MacBookPro 15, 4-core 2,2 GHz Intel Core i7, 16GB RAM, 1TB SSD running MacOSX 10.11: 53 seconds; • desktop computer, 8-core 3,4 GHz Intel i7, 16GB RAM, 2TB hard disk running Ubuntu: 45 seconds. The time is correlated with image size (0.94) and file size (0.89). 12
Page 12 of 17
4. Discussion and Conclusion
Ac ce p
te
d
M
an
us
cr
ip t
The identification of Ki67 hotspots in breast cancer biopsies is crucial for a correct diagnosis and treatment planning. The procedure of identification and quantification of the proliferation rate is normally made by the pathologist. The correct assessment of the Ki67 labeling index can be timeconsuming and is strictly related to the experience of the pathologist evaluating the slides. The AKHoD method efficiently and automatically provides the pathologist with suggestions regarding the localization of the hotspots along with some information such as the quantification of Ki67 labeling index as well as an estimation of the number of nuclei contained in the hotspot areas. The quantification of the Ki67 biomarker within the hotspot area is calculated as the ratio between the DAB and H components. The estimation of the number of cell nuclear sections within the hotspots contributes to increase the quality of the suggested hotspots by excluding small areas (< 500 nuclear sections) which can give raise to overestimated Ki67 values, as well as large areas (> 2000 nuclear sections) which can give rise to underestimated Ki67 values. These conditions satisfy the definition of Ki67 hotspot [1, 2], which makes the AKHoD method an important decision support tool for the pathologists. The use of BioFormats libraries contributes also to the flexibility of the AKHoD method, allowing it to read many of the WSI typical formats. The AKHoD method is applicable to WSI images acquired at different magnifications and provides also the flexibility to use different magnifications and samplings. The results are written following the OME-XML structure, making them readable from the most common WSI viewers. The hotspots suggested from the AKHoD system that were not identified by the pathologists may highlight areas that are challenging to be detected as hotspots. However, from a visual examination of the suggested hotspots is possible to see that some laboratory artifacts may have a color which intensity is similar to the positive DAB staining, making them difficult to discriminate automatically. Since the system provides only candidates for a further evaluation, the pathologist may easily discriminate between real hotspots and artifacts. In case that a quantitative evaluation of the suggested hotspots is performed automatically, it would be necessary to implement a method for the discrimination of tissue and artifacts. Some non-tumoral cells such as infiltrating lymphocites may also have a positivity to Ki67 and contribute to the definition of a suggested hotspot from AKHoD. A further 13
Page 13 of 17
Ac ce p
te
d
M
an
us
cr
ip t
examination from an expert pathologists may easily discriminate between hotspot suggestions. This aspect has also to be considered for an automated analysis by integrating methods for tumor tissue recognition such as Genie [16] that could be used to improve the quality of the suggested hotspots. In literature only few methods for the automatic recognition of the hotspots have been proposed and evaluated. A method that has been quantitatively evaluated was the ASH method [5]. Nielsen et al. [17] studied the application of the ASH method in the recognition of hotspots in melanoma biopsies. In particular such study found for the ASH method an average sensitivity of 67% and an average precision of 48%, while the AKHoD method showed an average sensitivity of 74.7% and an average precision of 75.5%. Even if the results of the AKHoD method are good in comparison with the hotspots detected from the expert pathologists, a more complete evaluation has to be performed taking into account also the follow up of the patients. In fact, guidelines over the Ki67 evaluation are still under discussion and recent studies approach the problem differently [18]. The average time performance is also sufficient and compatible with the daily routine of the pathologists. Possible extensions of the AKHoD method include the parallelization and an improved nuclear sections counting methodology to further improve both the performance and the quality of the suggested hotspot areas. The dataset tested in this work is assumed to have a certain constancy in the staining and preparation of the samples. As future work, more extensive validation will also be performed over slides from different laboratories and eventually slides acquired with different systems, considering also methods for the normalization of samples such as histogram normalization. 5. Conflict of Interest Statement The Authors declare no competing interests.
6. Acknowledgments
This work is partially founded by the EU FP7 program, AIDPATH project, grant number 612471.
14
Page 14 of 17
References
cr
ip t
[1] M. Dowsett, T. O. Nielsen, R. AHern, J. Bartlett, R. C. Coombes, J. Cuzick, M. Ellis, N. L. Henry, J. C. Hugh, T. Lively, et al., Assessment of Ki67 in breast cancer: recommendations from the International Ki67 in Breast Cancer working group, Journal of the National Cancer Institute.
an
us
[2] E. Luporsi, F. Andr´e, F. Spyratos, P.-M. Martin, J. Jacquemier, F. Penault-Llorca, N. Tubiana-Mathieu, B. Sigal-Zafrani, L. Arnould, A. Gompel, et al., Ki-67: level of evidence and methodological considerations for its role in the clinical management of breast cancer: analytical and critical review, Breast cancer research and treatment 132 (3) (2012) 895–915.
M
[3] J. Molin, K. S. Devan, K. W˚ ardell, C. Lundstr¨om, Feature-enhancing zoom to facilitate Ki-67 hot spot detection, in: SPIE Medical Imaging, International Society for Optics and Photonics, 2014, pp. 90410W– 90410W.
te
d
[4] V. J. Tuominen, S. Ruotoistenmaki, A. Viitanen, M. Jumppanen, J. Isola, Immunoratio: a publicly available web application for quantitative image analysis of estrogen receptor (ER), progesterone receptor (PR), and Ki-67, Breast Cancer Res 12 (4) (2010) R56.
Ac ce p
[5] H. Lu, T. G. Papathomas, D. van Zessen, I. Palli, R. R. de Krijger, P. J. van der Spek, W. N. Dinjens, A. P. Stubbs, Automated selection of hotspots (ASH): enhanced automated segmentation and adaptive step finding for Ki67 hotspot detection in adrenal cortical cancer, Diagnostic pathology 9 (1) (2014) 1–9. [6] M. K. Khan Niazi, M. M. Yearsley, X. Zhou, W. L. Frankel, M. N. Gurcan, Perceptual clustering for automatic hotspot detection from Ki67-stained neuroendocrine tumour images, Journal of microscopy 256 (3) (2014) 213–225. [7] B. Plancoulaine, A. Laurinaviciene, P. Herlin, J. Besusparis, R. Meskauskas, I. Baltrusaityte, Y. Iqbal, A. Laurinavicius, A methodology for comprehensive breast cancer Ki67 labeling index with intratumor heterogeneity appraisal based on hexagonal tiling of digital image analysis data, Virchows Archiv 467 (6) (2015) 711–722. 15
Page 15 of 17
ip t
[8] M. K. K. Niazi, E. Downs-Kelly, M. N. Gurcan, Hot spot detection for breast cancer in Ki-67 stained slides: image dependent filtering approach, in: SPIE Medical Imaging, International Society for Optics and Photonics, 2014, pp. 904106–904106.
us
cr
[9] G. Bueno, M. Fernandez-Carrobles, C. Womack, D. Segers, R. Ecker, S. Costello, V. Della Mea, E. Carasevici, M. Ilyas, G. Qiu, L. Gonzalez, A. Laurinavicius, G. Schaefer, M. Garcia-Rojo, O. Denizs, Digital pathology tools from and for academia and industry collaboration, In Proc. of 18th International Microscopy Congress (IMC2014).
an
[10] A. C. Ruifrok, D. A. Johnston, Quantification of histochemical staining by color deconvolution., Analytical and quantitative cytology and histology/the International Academy of Cytology [and] American Society of Cytology 23 (4) (2001) 291–299.
te
d
M
[11] I. G. Goldberg, C. Allan, J.-M. Burel, D. Creager, A. Falconi, H. Hochheiser, J. Johnston, J. Mellen, P. K. Sorger, J. R. Swedlow, The open microscopy environment (OME) data model and XML file: open tools for informatics and quantitative analysis in biological imaging, Genome biology 6 (5) (2005) 1.
Ac ce p
[12] M. Linkert, C. T. Rueden, C. Allan, J.-M. Burel, W. Moore, A. Patterson, B. Loranger, J. Moore, C. Neves, D. MacDonald, et al., Metadata matters: access to image data in the real world, The Journal of cell biology 189 (5) (2010) 777–782. [13] C. A. Schneider, W. S. Rasband, K. W. Eliceiri, et al., NIH image to imagej: 25 years of image analysis, Nat methods 9 (7) (2012) 671–675. [14] J. Schindelin, C. T. Rueden, M. C. Hiner, K. W. Eliceiri, The ImageJ ecosystem: An open platform for biomedical image analysis, Molecular reproduction and development 82 (7-8) (2015) 518–529. [15] J. Schindelin, I. Arganda-Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch, C. Rueden, S. Saalfeld, B. Schmid, et al., Fiji: an open-source platform for biological-image analysis, Nature methods 9 (7) (2012) 676–682.
16
Page 16 of 17
ip t
[16] A. Laurinavicius, A. Laurinaviciene, V. Ostapenko, D. Dasevicius, S. Jarmalaite, J. Lazutka, Immunohistochemistry profiles of breast ductal carcinoma: factor analysis of digital image analysis data, Diagnostic pathology 7 (1) (2012) 1.
us
cr
[17] P. S. Nielsen, R. Riber-Hansen, H. Schmidt, T. Steiniche, Automated quantification of proliferation with automated hot-spot selection in phosphohistone h3/mart1 dual-stained stage i/ii melanoma, Diagnostic pathology 11 (1) (2016) 1.
Ac ce p
te
d
M
an
[18] A. Laurinavicius, B. Plancoulaine, A. Rasmusson, J. Besusparis, R. Augulis, R. Meskauskas, P. Herlin, A. Laurinaviciene, A. A. A. Muftah, I. Miligy, et al., Bimodality of intratumor Ki67 expression is an independent prognostic factor of overall survival in patients with invasive breast carcinoma, Virchows Archiv 468 (4) (2016) 493–502.
17
Page 17 of 17
S0895-6111(17)30036-8 http://dx.doi.org/doi:10.1016/j.compmedimag.2017.04.005 CMIG 1506
To appear in:
Computerized Medical Imaging and Graphics
Received date: Revised date: Accepted date:
19-7-2016 18-4-2017 23-4-2017
Please cite this article as: David Pilutti, Vincenzo Della Mea, Enrico Pegolo, Francesco La Marra, Fulvio Antoniazzi, Carla Di Loreto, An adaptive positivity thresholding method for automated Ki67 hotspot detection (AKHoD) in breast cancer biopsies, (2017), http://dx.doi.org/10.1016/j.compmedimag.2017.04.005 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ip t
An adaptive positivity thresholding method for automated Ki67 hotspot detection (AKHoD) in breast cancer biopsies
cr
David Piluttia , Vincenzo Della Meaa , Enrico Pegolob , Francesco La Marrac , Fulvio Antoniazzic , Carla Di Loretoc a
an
us
Department of Mathematics, Computer Science, and Physics, University of Udine, Udine, Italy b Institute of Pathology, University Hospital ”Santa Maria della Misericordia”, Udine, Italy c Department of Medical and Biological Science, University of Udine, Udine, Italy
M
Abstract
Ac ce p
te
d
The proliferative activity of breast cancer tissue can be estimated using the Ki67 biomarker. The percentage of positivity of such biomarker is correlated with proliferation and consequently with the prognosis of a breast tumor. Ki67 marked tissue samples are analyzed by an experienced pathologist who identifies the most active areas of tumor cell proliferation called hotspots, and estimates the positivity of each case. A method for the Automated Ki67 Hotspot Detection (AKHoD) is presented in this work. The main objective of the AKHoD method is to automatically and efficiently provide the pathologist with suggestions about Ki67 hotspot areas as a decision support. The input of AKHoD is a digital slide that is divided in tiles. For each tile, AKHoD provides a rough estimate of positivity and cellularity, summarized in very low resolution positivity and cellularity images. In a second step, an adaptive thresholding is applied to such positivity image to identify the most positive connected and convex areas, within cellularity limits set by current guidelines (that is, 500-2000). The method has been preliminarily validated on 50 digital slides for which three expert pathologists provided gold standard hotspots. 82% of the gold standard hotspots have been successfully recognized by the system, spending an average of 54 seconds per slide. While further validation is needed taking into account also patients followup, this first experimentation suggests that the proposed method could be adequate for supporting the pathologist in Preprint submitted to Computerized Medical Imaging and Graphics
April 18, 2017
Page 1 of 17
hotspot detection.
ip t
Keywords: Digital Pathology, Ki67, Automated Hotspot Detection, Breast Cancer, Image Analysis
cr
1. Background
Ac ce p
te
d
M
an
us
The identification of hotspots as well as the quantification of immunohistochemical Ki67 positive staining are of critical importance for the prognosis and the treatment planning of breast cancer. A high percentage of Ki67 positivity within a biopsy indicates an higher proliferation rate and consequently the aggressiveness of a breast tumor. The proliferation rate is normally visually estimated by the pathologist over an hotspot area that should include from 500 to 2000 cells [1, 2]. This makes the identification and quantification of the hotspot areas as well as the calculation of the proliferation rate complicated and time consuming. To address such tasks, different approaches have been used. A semiautomated method for the identification of Ki67 hotspots has been proposed [3]. This method reduces the zooming operations that a pathologist must perform to select a hotspot area by enlarging positive nuclei at low magnification to make them easier to identify. ImmunoRatio is an open source web based application which performs a calculation of positively stained nuclear area by applying color deconvolution and adaptive thresholding [4]. The input is represented by stained images acquired with digital cameras and compressed in common formats such as JPEG, JPEG2000, TIFF, BMP, and PNG. The execution of ImmunoRatio generates a pseudo-colored segmented image which can be successively analyzed by an expert. Fully automated methods for the identification of Ki67 hotspot areas have also been developed. A method for the detection of Ki67 hotspot areas is the Automated Selection of Hotspots (ASH) [5], a Linux desktop application. The input of the ASH method is a Hamamatsu NDPI image which is splitted into smaller images called image blocks. Each image block is converted in a common image format such as TIFF or JPEG and analyzed using ImmunoRatio [4] to rank the top 10 image blocks, which represent potential hotspots. Clustering methodologies have also been used to address the hotspot detection in Ki67 stained neuroendocrine tumor images [6]. In such clustering framework the positive cell nuclear sections are segmented using minimal 2
Page 2 of 17
te
d
M
an
us
cr
ip t
graph cuts. The resulting segmentation is then used to detect the hotspots using non linear transformation for the clustering, allowing inclusions of positive nuclear sections in a cluster even in the case of sparse nuclei. A learning algorithm is combined with the clustering framework to simulate the human perception of an hotspot. An alternative method [7] detects Ki67 hotspots using hexagonal tiling and implementing the concept of Pareto hotspot, which represents the upper quintile of the biomarker expression in the tissue, providing also a quantitative analysis over the detected hotspots. The hotspot detection for Ki67 stained slides of breast cancer has also been posed as an image filtering problem [8]. This approach considers all the pixels corresponding to positive stained nuclei as a surrogate representation of a nuclear section. A lowpass filter is then applied to such image to obtain an approximated density map where the local maxima contribute to the identification of the hotspots in the image. In this paper a new efficient and robust method for the Automated Ki67 Hotspot Detection (AKHoD) is proposed with the objective to provide the pathologists with a fast and robust method for decision support applied to the analysis of breast cancer biopsies. The main goal of the AKHoD method is to automatically identify hotspot areas over Ki67 stained breast cancer biopsies images and provide the pathologist with visual suggestions. An approximated estimation of the proliferation rate as well as an estimation of cell nuclear section is also calculated within the suggested hotspot areas.
Ac ce p
2. Materials and Methods
The digital slides used for this study have been previously acquired as material for the AIDPATH breast cancer database [9]. On 50 invasive breast cancer biopsy samples immunohistochemistry was performed using the Ki67/MIB1 antibody, stained with Diaminobenzidine (DAB) and counterstained with Hematoxylin (H), resulting in H/DAB staining. The samples have been acquired using an Aperio CS (Leica Biosystems, Nussloch, Germany) slide scanner at 40x, with a resolution of 0.2491µm per pixel. Slides were saved in TIFF using the proprietary Leica SVS format. Input Whole Slide Images (WSI) are divided in tiles, whose colors are deconvolved using the method of Ruifrok [10] using the standard Hemathoxylin/Diaminobenzidine (H/DAB) deconvolution matrix and assuming a certain constancy in sample staining and preparation. This results into two
3
Page 3 of 17
Ac ce p
te
d
M
an
us
cr
ip t
deconvolved images which express the H and the DAB color component, respectively. The presence of Ki67 is quantified by calculating a ratio between the DAB and the H component. The number of tiles varies depending on the magnification level selected for the analysis. This results in a grouping of tiles that reduces the computation time needed for the analysis. The result is a positivity image in which each pixel corresponds to a tile in the input image and the intensities are the calculated DAB over H average positivity ratios. A dynamic color threshold is applied to such image to highlight potential hotspot areas. The optimization procedure starts from the maximum normalized density value used as color threshold that is iteratively diminished until the found hotspot areas have a sufficient size to contain an adequate number of nuclear sections. The stop criterion for iterations has been set to 2000 nuclear sections. The resulting hotspot areas are then written in an OME-XML [11] file which can be read by a medical image viewer such as Aperio ImageScope. To provide the pathologist with more significative suggestions, the three hotspot areas with highest DAB/H ratio are highlighted using a different color. The DAB/H ratio is calculated within each hotspot area and the result is also provided in the output. Hotspot areas have been identified and annotated over the 50 considered cases by three experienced pathologists. The manual annotations have been compared with the results from the AKHoD method. A percent overlap between hotspot areas from the experts and from AKHoD has been calculated. This gives a quantitative measure of the quality of the results obtained with the AKHoD method.
2.1. Positivity Quantification and Cellularity Estimation The pixels Hp and DABp corresponding to the H and DAB channel, respectively, are counted and the ratio (DABp /Hp )∗100 is calculated, resulting in a percent estimation of the DAB staining within a tile. This contributes to create a positivity image IP OS in which each pixel corresponds to a tile and the intensities correspond to the estimated DAB densities. In Fig. 1 an example of a WSI image stained with H/DAB and the corresponding positivity image obtained is shown, where a high intensity corresponds to a high DAB staining. The quantification of the DAB staining has to be integrated with an estimation of the number of cell nuclear sections within a tile to provide a more reliable result for determining the size of an hotspot. The average 4
Page 4 of 17
te
d
M
an
us
cr
ip t
radius for a breast tumor nuclear section is assumed to be 10µm. This value is used to statistically estimate the number of DAB stained nuclear sections by dividing the area corresponding to the DAB staining with the average nuclear area. This results in a cellularity image INDAB where the intensity of each pixel represents the number of the counted nuclear sections. A similar estimation is performed for the H staining assuming an average size of the cell nuclear section of 7µm, with the exclusion of double colored areas to obtain the image INH . In both estimations the background area is also excluded. In Fig. 2 there are examples of the nuclei density estimation of DAB and H staining components.
(b) positivity image
Ac ce p
(a) WSI image
Figure 1: Example of a WSI image of a breast tumor biopsy stained with H/DAB and its corresponding positivity image. In (b) each pixel correspond to a tile in (a). The intensities in (b) represent the DAB/H density, where an high intensity corresponds to a high DAB staining.
2.2. Hotspots Estimation The estimation of the hotspots is performed over IP OS with the support of INDAB and INH . The positivity image IP OS is thresholded starting from the threshold td = 100 which is the maximum value possible. The thresholded image is used to segment positive regions, calculate the relative areas with the average positivity percent value. This results in a table in which each line represents a different segmented region of the positivity image, providing 5
Page 5 of 17
ip t cr us (b) H nuclei
an
(a) DAB nuclei
M
Figure 2: Example of cellularity images. In (a) is an estimated DAB stained cellularity image. In (b) is an estimated H stained cellularity image with the exclusion of the double stained pixels. The intensity represents the number of the counted nuclear sections.
Ac ce p
te
d
also information about the position and the size of each region. The position and the size of the resulting regions are then also localized in INDAB and INH to perform a nuclear section count within the selected area. The regions containing a total number of nuclear sections nc below 500 are excluded from the results table. The AKHoD method iterates by lowering td and consequently progressively analyzing bigger parts of IP OS until one candidate hotspot area contains more than 2000 nuclear sections or reaches the maximum number of iterations k = 100. In Fig. 3 there is an example of a positivity image IP OS of Fig. 1b to which three different thresholds td are applied, highlighting different areas of IP OS . At the end of the iterations, the resulting candidate hotspot areas are classified with respect of their average DAB positivity value and written in the OME-XML output file. In Fig. 4 there is an overview of the AKHoD method with the separation between the preparatory phase of defining ID , INDAB , and INH and the iterations with different td. If AKHoD identifies multiple candidate hotspot areas, the three areas of higher DAB positivity expression are highlighted using a red color. Other candidates are highlighted using the green color to differentiate them from the areas of higher DAB positivity expression.
6
Page 6 of 17
ip t (b) positivity td = 90
image,
cr
image,
(c) positivity td = 80
image,
us
(a) positivity td = 100
Ac ce p
te
d
M
an
Figure 3: An example of a positivity image to which are applied different thresholds td resulting in the highlight of progressively bigger areas of the image when lowering td.
Figure 4: Overview of AKHoD method. The positivity image ID is calculated in the preparatory phase together with INDAB and INH . The iterations are performed decreasing the value of td until the match between the estimations of DAB and the number of cell nuclear sections satisfy the definition of an hotspot.
7
Page 7 of 17
Ac ce p
te
d
M
an
us
cr
ip t
2.3. Validation Three experienced pathologists reviewed the Ki67 stained glass slides that were previously digitally acquired and marked the hotspot areas. These marked areas were then reported on the digital slides as annotations using the Aperio ImageScope software. Annotated digital slides provided a gold standard for evaluation. The validation has been performed semi-quantitatively by calculating the superimposition between the suggested candidate hotspots and the gold standard. A gold standard hotspot has been considered recognized by the system if, at visual analysis, there is significant superimposition between both. In fact, since both gold standard and calculated hotspots are rough shapes, having full correspondence is almost impossible. A gold standard hotspot has been considered ”missed” if there is no superimposition with calculated hotspots. A calculated hotspot is considered ”added” if there is no corresponding gold standard hotspot. The number of recognized, missed, and added hotspots is calculated for each slide. Since one aim of the present work is to provide a fast system for decision support to the pathologist, the AKHoD method has been tested on different series (40x, 10x, 2.5x) included in the slide file, using also different tile sizes (60, 120, 240). The best compromise between recognition capabilities and speed has been then chosen. After the selection of one set of parameters, tests have been performed on 3 different machines, recording the time needed for execution on each slide to evaluate the average time performance. Execution time has been compared with image and file size using Pearson correlation.
2.4. Implementation The AKHoD method has been implemented in Java using the BioFormats library [12] to read WSI images. The tiles have been deconvolved using the color deconvolution method of Ruifrok [10] implemented in ImageJ. The functions of thresholding and segmentation of positive regions with ParticleAnalyzer have been integrated into the main code by using ImageJ as Java library [13, 14, 15]. The output XML files containing the hotspot suggestions have been created using the OME-XML structure [11]. The testing has been performed on three different machines, two workstations and a server.
8
Page 8 of 17
3. Results
Ac ce p
te
d
M
an
us
cr
ip t
The patient datasets used in this work have been previously acquired in the context of clinical studies which were approved by the institutional ethical committee, and informed consent was obtained from all patients. All the datasets have been also anonymized before testing. All the datasets have been provided with an indication of gold standard hotspots made by experienced pathologists. In Fig. 5 there are the comparative results of gold standard and AKHoD results in a case with positivity of 30% resulting in a good match between gold standard hotspots and AKHoD suggestions.
(a) Gold standard
(b) AKHoD suggestions
Figure 5: Comparative examples between gold standard hotspots (a) and the suggested areas from AKHoD (b) over a case of etimated positivity of 30%. All the hotspots area from the gold standard have been detected also from the AKHoD method.
In Fig. 6, instead, the AKHoD method has been tested on a case with positivity of 7%. It can be seen that both gold standard hotspot have been 9
Page 9 of 17
te
d
M
an
us
cr
ip t
also detected from the AKHoD method, which adds another area that have to be further validated by the pathologist.
(b) AKHoD suggestions
Ac ce p
(a) Gold standard
Figure 6: Comparative examples between gold standard hotspots (a) and the suggested areas from AKHoD (b) over a case of estimated positivity of 7%. Both of the hotspots areas from the gold standard have been detected also from the AKHoD method. A third hotspot area has been detected only from the AKHoD method, providing information that has to be further validated from the expert pathologist.
Finally, in Fig. 7 (65% positivity) one of the suggestions from the AKHoD method matched a gold standard hotspot, while the second suggestion resulted in a merge of two adjacent gold standard hotspot areas. The preliminary optimization of the algorithm parameters allowed to choose the series 2 (2.5x) with a tile size of 60 pixels as the best compromise between recognition performance and speed. In fact, the best one was series 0 (40x) with tile size 240 pixels, that recognized one hotspot more than the selected one, but with an average time needed for recognition 15 times higher. 10
Page 10 of 17
ip t cr us an M
(a) Gold standard
(b) AKHoD suggestions
te
d
Figure 7: Comparative examples between gold standard hotspots (a) and the suggested areas from AKHoD (b) over a case of estimated positivity of 65%. In this case one hotspot area has been correctly detected from the AKHoD method while the other is the result of the merge between two adjacent hotspot areas in the gold standard image.
Ac ce p
The AKHoD method has been run on the above mentioned slides, correctly recognizing 59 out of 72 gold standard hotspots (81.94%) on which all the three experts agree. Taking into consideration the hotspots on which at least two pathologists agree, the AKHoD system correctly recognized 84 hotspots out of 107 (78.50%). Finally, considering the hotspots on which at least one pathologist agree, the AKHoD system correctly recognized 118 hotspots out of 172 (60.61%). In some cases, system-calculated hotspots were overlapping with two adjacent gold standard hotspots. Table 1 shows details about correctly recognized hotspots and compare them with the agreement of the pathologists. Some statistical indexes such as Dice, Jaccard, sensitivity, and precision have also been calculated by comparing the number of hotspots recognized by the system and the hotspot areas provided by the pathologists. The overlapping hotspots have been considered as the True Positive (TP) values, while the False Positive (FP) are the hotspots recognized by the AKHoD
11
Page 11 of 17
Pathologists hotspots % 72 81.94 107 78.50 172 60.61
ip t
Agreement AKHoD n. Pathologists correct hotspots 3 59 2 84 1 118
us
cr
Table 1: Table presenting the number of the correctly recognized hotspots from the AKHoD system in comparison with the hotspots identified by the pathologists over the whole dataset. The rows provide the number of hotspots recognized by the AKHoD system over the number of hotspots on which at least three, two, and one patologists agree, respectively.
Pathologist 2 0.764 0.618 0.752 0.776
te
d
Index Pathologist 1 Dice 0.770 Jaccard 0.628 Sensitivity 0.784 Precision 0.757
M
an
system but not from the pathologists, and finally the False Negative (FN) are the hotspots that have been recognized by the pathologists but not by the system. As shown in Table 2, the average Dice coefficient resulted in 0.751 and the Jaccard in 0.602 with a sensitivity of 0.747 and a precision of 0.755. Pathologist 3 0.720 0.562 0.706 0.733
Average 0.751 0.602 0.747 0.755
Ac ce p
Table 2: Table describing the performance of the AKHoD method in comparison with the results obtained from the experts.
Although the prototype implementation was not yet optimized, the AKHoD method has been tested in 3 different systems to preliminary evaluate its average time performance: • MacBook 13, 2-core 2,8 GHz Intel Core i7, 16GB RAM, 512MB SSD running Windows 7: 84 seconds; • MacBookPro 15, 4-core 2,2 GHz Intel Core i7, 16GB RAM, 1TB SSD running MacOSX 10.11: 53 seconds; • desktop computer, 8-core 3,4 GHz Intel i7, 16GB RAM, 2TB hard disk running Ubuntu: 45 seconds. The time is correlated with image size (0.94) and file size (0.89). 12
Page 12 of 17
4. Discussion and Conclusion
Ac ce p
te
d
M
an
us
cr
ip t
The identification of Ki67 hotspots in breast cancer biopsies is crucial for a correct diagnosis and treatment planning. The procedure of identification and quantification of the proliferation rate is normally made by the pathologist. The correct assessment of the Ki67 labeling index can be timeconsuming and is strictly related to the experience of the pathologist evaluating the slides. The AKHoD method efficiently and automatically provides the pathologist with suggestions regarding the localization of the hotspots along with some information such as the quantification of Ki67 labeling index as well as an estimation of the number of nuclei contained in the hotspot areas. The quantification of the Ki67 biomarker within the hotspot area is calculated as the ratio between the DAB and H components. The estimation of the number of cell nuclear sections within the hotspots contributes to increase the quality of the suggested hotspots by excluding small areas (< 500 nuclear sections) which can give raise to overestimated Ki67 values, as well as large areas (> 2000 nuclear sections) which can give rise to underestimated Ki67 values. These conditions satisfy the definition of Ki67 hotspot [1, 2], which makes the AKHoD method an important decision support tool for the pathologists. The use of BioFormats libraries contributes also to the flexibility of the AKHoD method, allowing it to read many of the WSI typical formats. The AKHoD method is applicable to WSI images acquired at different magnifications and provides also the flexibility to use different magnifications and samplings. The results are written following the OME-XML structure, making them readable from the most common WSI viewers. The hotspots suggested from the AKHoD system that were not identified by the pathologists may highlight areas that are challenging to be detected as hotspots. However, from a visual examination of the suggested hotspots is possible to see that some laboratory artifacts may have a color which intensity is similar to the positive DAB staining, making them difficult to discriminate automatically. Since the system provides only candidates for a further evaluation, the pathologist may easily discriminate between real hotspots and artifacts. In case that a quantitative evaluation of the suggested hotspots is performed automatically, it would be necessary to implement a method for the discrimination of tissue and artifacts. Some non-tumoral cells such as infiltrating lymphocites may also have a positivity to Ki67 and contribute to the definition of a suggested hotspot from AKHoD. A further 13
Page 13 of 17
Ac ce p
te
d
M
an
us
cr
ip t
examination from an expert pathologists may easily discriminate between hotspot suggestions. This aspect has also to be considered for an automated analysis by integrating methods for tumor tissue recognition such as Genie [16] that could be used to improve the quality of the suggested hotspots. In literature only few methods for the automatic recognition of the hotspots have been proposed and evaluated. A method that has been quantitatively evaluated was the ASH method [5]. Nielsen et al. [17] studied the application of the ASH method in the recognition of hotspots in melanoma biopsies. In particular such study found for the ASH method an average sensitivity of 67% and an average precision of 48%, while the AKHoD method showed an average sensitivity of 74.7% and an average precision of 75.5%. Even if the results of the AKHoD method are good in comparison with the hotspots detected from the expert pathologists, a more complete evaluation has to be performed taking into account also the follow up of the patients. In fact, guidelines over the Ki67 evaluation are still under discussion and recent studies approach the problem differently [18]. The average time performance is also sufficient and compatible with the daily routine of the pathologists. Possible extensions of the AKHoD method include the parallelization and an improved nuclear sections counting methodology to further improve both the performance and the quality of the suggested hotspot areas. The dataset tested in this work is assumed to have a certain constancy in the staining and preparation of the samples. As future work, more extensive validation will also be performed over slides from different laboratories and eventually slides acquired with different systems, considering also methods for the normalization of samples such as histogram normalization. 5. Conflict of Interest Statement The Authors declare no competing interests.
6. Acknowledgments
This work is partially founded by the EU FP7 program, AIDPATH project, grant number 612471.
14
Page 14 of 17
References
cr
ip t
[1] M. Dowsett, T. O. Nielsen, R. AHern, J. Bartlett, R. C. Coombes, J. Cuzick, M. Ellis, N. L. Henry, J. C. Hugh, T. Lively, et al., Assessment of Ki67 in breast cancer: recommendations from the International Ki67 in Breast Cancer working group, Journal of the National Cancer Institute.
an
us
[2] E. Luporsi, F. Andr´e, F. Spyratos, P.-M. Martin, J. Jacquemier, F. Penault-Llorca, N. Tubiana-Mathieu, B. Sigal-Zafrani, L. Arnould, A. Gompel, et al., Ki-67: level of evidence and methodological considerations for its role in the clinical management of breast cancer: analytical and critical review, Breast cancer research and treatment 132 (3) (2012) 895–915.
M
[3] J. Molin, K. S. Devan, K. W˚ ardell, C. Lundstr¨om, Feature-enhancing zoom to facilitate Ki-67 hot spot detection, in: SPIE Medical Imaging, International Society for Optics and Photonics, 2014, pp. 90410W– 90410W.
te
d
[4] V. J. Tuominen, S. Ruotoistenmaki, A. Viitanen, M. Jumppanen, J. Isola, Immunoratio: a publicly available web application for quantitative image analysis of estrogen receptor (ER), progesterone receptor (PR), and Ki-67, Breast Cancer Res 12 (4) (2010) R56.
Ac ce p
[5] H. Lu, T. G. Papathomas, D. van Zessen, I. Palli, R. R. de Krijger, P. J. van der Spek, W. N. Dinjens, A. P. Stubbs, Automated selection of hotspots (ASH): enhanced automated segmentation and adaptive step finding for Ki67 hotspot detection in adrenal cortical cancer, Diagnostic pathology 9 (1) (2014) 1–9. [6] M. K. Khan Niazi, M. M. Yearsley, X. Zhou, W. L. Frankel, M. N. Gurcan, Perceptual clustering for automatic hotspot detection from Ki67-stained neuroendocrine tumour images, Journal of microscopy 256 (3) (2014) 213–225. [7] B. Plancoulaine, A. Laurinaviciene, P. Herlin, J. Besusparis, R. Meskauskas, I. Baltrusaityte, Y. Iqbal, A. Laurinavicius, A methodology for comprehensive breast cancer Ki67 labeling index with intratumor heterogeneity appraisal based on hexagonal tiling of digital image analysis data, Virchows Archiv 467 (6) (2015) 711–722. 15
Page 15 of 17
ip t
[8] M. K. K. Niazi, E. Downs-Kelly, M. N. Gurcan, Hot spot detection for breast cancer in Ki-67 stained slides: image dependent filtering approach, in: SPIE Medical Imaging, International Society for Optics and Photonics, 2014, pp. 904106–904106.
us
cr
[9] G. Bueno, M. Fernandez-Carrobles, C. Womack, D. Segers, R. Ecker, S. Costello, V. Della Mea, E. Carasevici, M. Ilyas, G. Qiu, L. Gonzalez, A. Laurinavicius, G. Schaefer, M. Garcia-Rojo, O. Denizs, Digital pathology tools from and for academia and industry collaboration, In Proc. of 18th International Microscopy Congress (IMC2014).
an
[10] A. C. Ruifrok, D. A. Johnston, Quantification of histochemical staining by color deconvolution., Analytical and quantitative cytology and histology/the International Academy of Cytology [and] American Society of Cytology 23 (4) (2001) 291–299.
te
d
M
[11] I. G. Goldberg, C. Allan, J.-M. Burel, D. Creager, A. Falconi, H. Hochheiser, J. Johnston, J. Mellen, P. K. Sorger, J. R. Swedlow, The open microscopy environment (OME) data model and XML file: open tools for informatics and quantitative analysis in biological imaging, Genome biology 6 (5) (2005) 1.
Ac ce p
[12] M. Linkert, C. T. Rueden, C. Allan, J.-M. Burel, W. Moore, A. Patterson, B. Loranger, J. Moore, C. Neves, D. MacDonald, et al., Metadata matters: access to image data in the real world, The Journal of cell biology 189 (5) (2010) 777–782. [13] C. A. Schneider, W. S. Rasband, K. W. Eliceiri, et al., NIH image to imagej: 25 years of image analysis, Nat methods 9 (7) (2012) 671–675. [14] J. Schindelin, C. T. Rueden, M. C. Hiner, K. W. Eliceiri, The ImageJ ecosystem: An open platform for biomedical image analysis, Molecular reproduction and development 82 (7-8) (2015) 518–529. [15] J. Schindelin, I. Arganda-Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch, C. Rueden, S. Saalfeld, B. Schmid, et al., Fiji: an open-source platform for biological-image analysis, Nature methods 9 (7) (2012) 676–682.
16
Page 16 of 17
ip t
[16] A. Laurinavicius, A. Laurinaviciene, V. Ostapenko, D. Dasevicius, S. Jarmalaite, J. Lazutka, Immunohistochemistry profiles of breast ductal carcinoma: factor analysis of digital image analysis data, Diagnostic pathology 7 (1) (2012) 1.
us
cr
[17] P. S. Nielsen, R. Riber-Hansen, H. Schmidt, T. Steiniche, Automated quantification of proliferation with automated hot-spot selection in phosphohistone h3/mart1 dual-stained stage i/ii melanoma, Diagnostic pathology 11 (1) (2016) 1.
Ac ce p
te
d
M
an
[18] A. Laurinavicius, B. Plancoulaine, A. Rasmusson, J. Besusparis, R. Augulis, R. Meskauskas, P. Herlin, A. Laurinaviciene, A. A. A. Muftah, I. Miligy, et al., Bimodality of intratumor Ki67 expression is an independent prognostic factor of overall survival in patients with invasive breast carcinoma, Virchows Archiv 468 (4) (2016) 493–502.
17
Page 17 of 17