- Email: [email protected]
Dendritic Cell Recognition in Computer Aided System for Cancer Immunotherapy
Available online at www.sciencedirect.com
ScienceDirect Procedia Computer Science 105 (2017) 177 – 182
2016 IEEE International Symposium on Robotics and Intelligent Sensors, IRIS 2016, 17-20 December 2016, Tokyo, Japan
Dendritic Cell Recognition in Computer Aided System for Cancer Immunotherapy Anis Azwani Muhd Suberiª, Wan Nurshazwani Wan Zakaria*ª, Razali Tomariª ªAdvance Mechatronic Research Group (ADMIRE), Faculty of Electrical and ElectronicEngineering, Universiti Tun Hussein Onn Malaysia, Parit Raja Batu Pahat 86400, Malaysia
Abstract Immunotherapy is an entirely advanced class of cancer treatment which has been highly active and exciting field in clinical therapeutics. In numerous procedures, cancer immunotherapy demands a laborious practice to recognise and count Dendritic Cells (DCs) in the vaccine preparation. Conventionally, the laser-based technology that provides a rapid analysis such as Flow Cytometry can affect the DCs viability as the staining procedure is involved. Another highly promising method which is Phase Contrast Microscopy (PCM) involves experienced pathologists to visually examine the respective microscopy images. In facts, PCM confronts complex issues regarding to imaging artifacts which can deteriorate the recognition process. As the DCs counting is crucial in cancer treatment procedures, this paper proposes a pioneering system called CasDC which implement an image processing algorithm to recognise and count DCs with a label-free method. The aim of developing this system is to establish a reliable and time saving-tool as a second reader in the clinical practice. In the meantime, the treatment procedure can be administered and therefore, improve the patient’s survival rate. Our proposed system has an enormous potential towards helping Cancer Research Institute in which the system offers rapid and high throughput cancer immunotherapy vaccine preparation and automated live cell investigation. © 2017 byby Elsevier B.V. This is an open access article under the CC BY-NC-ND license © 2016 The TheAuthors. Authors.Published Published Elsevier B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors (IRIS2016). 2016). Sensors(IRIS Keywords: Dendritic cells; Cancer immunotherapy; Image processing; Pattern recognition; Phase contrast microsocopy; Computer vision
*
Corresponding author. Tel.: +6-07-453-7000; fax: +6-07-453-6060. E-mail address: [email protected]
1877-0509 © 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors(IRIS 2016). doi:10.1016/j.procs.2017.01.201
178
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
1.Introduction Most of the established therapies such as surgery, chemotherapy and radiation are widely used to treat cancer patients. Surgery [1] is often considered as the first therapy option to remove the tumour. Nevertheless, there is a tendency of removing only part of the tumour. Following that, a combination of surgery with radiation or chemotherapy is typically used in patient to kill the cancer cells [1]. However, both radiation and chemotherapy can introduce non-favourable outcomes towards the patient, such as serious bleeding, lack of energy and experience depression. Recently, immunotherapy has been widely explored and introduced as an advanced approach to boost the immune system to fight cancer [1–3]. Immunotherapy employs and activates the Dendritic Cells (DCs) to identify the cancer tissue in the normal body cells. DCs have a key role for activation of T- and B-cell immunity due to their superior ability to function as antigen-presenting cells (APCs) [2,4]. They can be generated in vitro from Peripheral Blood Mononuclear Cells (PBMCs) [4]. In DCs vaccine cell preparation, the identification of DCs from PBMCs sample is crucial before the cell can be stimulated. Conventional Flow Cytometry provides an effective recognition of labelled DCs using fluorescent dyes that may cause phototoxic damage to the DCs. Recent advances in cellular imaging have facilitated investigation of the unstained living cell using Phase Contrast Microscopy (PCM). Even though PCM is a label-free imaging modality, the identification purpose becomes challenging as the image is constituted with a variation of imaging artifacts such as halo region, low contrast and overlapping DCs [5]. Such procedure is laborious, time-consuming and very dependent on the expert’s skill to identify and count DCs which can introduce human errors. Previous researches introduce Computer Aided Diagnosis (CAD) [6,7] which is an active area of research and development in medical imaging and diagnostics. A system called ImPatho [8], has been developed for disease identification. This system helps to provide a proper clinical guideline towards any disease, such as RBC disorder. Although most clinical applications are developed for cancer detection, it can be expected that the DCs counting in immunotherapy will be an important component of cancer treatment. In this paper, the challenges and complex issues faced by Cancer Research Institute are addressed. Our proposed system, CasDC (Computer Aided System for Dendritic Cells), is developed to identify and count DCs in the sample. The main goal of CasDC system is to assist pathologists and other clinical practitioners in initiating the vaccine preparation for cancer immunotherapy treatment. The details of the proposed framework are described in Section 2, system overview, followed by the results and discussion in Section 3. Finally, the conclusion is summarized in Section 4. 2.System Overview In this section, the CasDC workflow is presented in Fig. 1. It consists of two stages; a) Image acquisition and b) image processing in CasDC. The system is developed, trained and tested using Matlab R2015a platform and runs on 2.5GHz i3-312M processor. The image dataset is provided by Cancer Research Malaysia (CRM) which composed of debris (black dots), T-cells (round shape) and DCs (red circle) as shown in Fig. 2. The digital images of PBMCs are captured by Nikon DS-Fi2 camera which attached to Olympus CK40 microscope with 100x and 200x magnification. All the images have variations of light-variant environments. PBMCs sample
Microscope Image acquisition
Image normalization
Pre-processing
Classification
Feature Extraction
Segmentation
Blood smear
Fig. 1. CasDC workflow
Fig. 2 PCM image sample
179
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
2.1.Workflow of CasDC Image Analysis Fig. 3 demonstrates the workflow of CasDC in DCs recognition and counting. Fig. 3(a) and (b) show the sample images are prepared for processing and counting process. A cell library that composed of DCs and other cells (T-cell and debris) is provided where it acts as a template which will be compared to the testing image to detect the most approximate cell signatures. During the matching process, One Dimensional (1D) Fourier Descriptors computation is used to compare the template with the testing image, which can enable rapid and efficient template matching to identified, mark and count DCs as shown in Fig. 3(c) and (d). (a)
Start Insert
(b)
Templates and testing images
Pre-processing and Segmentation
(d) (c)
Template matching based on 1D Fourier Descriptors
DCs recognition, marking and counting
Sample image
Display count result
End
Fig. 3. Workflow of CasDC
2.2.Algorithm for Dendritic Cells (DCs) Detection The key element of CasDC system is an image processing algorithm for recognising DCs as shown in Fig. 4. As an initial step, image normalization is applied to each testing image to reduce the illumination effect. Based on the previous work , a combination of Local Contrast Threshold (LCT) with halo removal [9] is applied as the PCM images are contaminated with imaging artifacts. However, the image contains overlapping and clumping cells which can deteriorate the classification process. To overcome these problems, a hybrid of low and high sigma in Gaussian kernel filtering through logical operator AND with Local Adaptive Threshold (H-GLAT) [10] have been applied in which this method address the imaging artifacts issues in overlapping and clumping cells as shown in Fig. 4(b). Subsequently, the segmentation masks are post-processed by several morphology operations and Canny edge to segregate the entire cell shapes in the image as shown in Fig. 4(c) and (d) respectively. The Fourier Descriptors (FDs) have been applied for recognition purposes and this method involves One-Dimensional (1D) version to represent the contour in the form of shape signatures. Towards the end, the classification stage is performed by using template matching [5]. The identified DCs are marked with red colour as shown in Fig. 4(e).
a)
b)
c)
d)
e)
Fig. 4. a) Original image b) Pre-processing (H-GLAT) c) Morphological operators d) Canny edge and e) DCs detection
180
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
2.3.Graphical User Interface (GUI) Overview Graphical user interface (GUI) has been developed for CasDC to facilitate the pathologist to identify and count DC as shown in Fig. 5. This automated system provides user-friendly features to the user in which the sample images can be processed into two modes; 1) single and 2) batch. Both modes can be operated for 100x and 200x magnification. Once the mode is selected, the user needs to press ‘Start Analysis’ button which will load image, process and extract the cell shape. The system continuously generates Fourier Descriptors (FDs) through the entire image to find the approximate similarity between the DCs template and the sample image. For a single image, the user needs to save the result manually using ‘Save Result’ button. On the other hand, for batch file mode, the system will continuously scan, mark and count DCs in the entire file of images and automatically save the results once the process is completed as shown in Fig. 6.
Fig. 5. CasDC: DCs recognition and counting GUI
Fig. 6. Result of batch mode
3.Results and Discussion Two image dataset in the JPEG picture format are used for evaluating system performance. The first and second dataset are composed of 1877x1408 (n=10) and 2560x1920 (n=135) pixels images respectively. Both image dataset consists with variations of image colour, light environments and shapes of DCs. Table 1 show some of the comparison of DCs counting between CasDC and manual with processing time. The identified DCs of CasDC and manual are marked with red colour and black arrow respectively. Meanwhile, the average processing time to run the system for each image is 2.229 seconds. Next, the quantitative measurement is performed based on a) precision, b) recall and c) accuracy as shown in Equation 1 [11]. Precision refers to the exact amount of DCs and recall presents the exact number of DCs from the input image. Meanwhile, accuracy measures the overall system performance in classifying DCs. Pr ecision =
TP TP TP + TN ; Re call = ; Accuracy = TP + FP TP + FN TP + FP + TN + FN
(1)
The lowest performance in term of precision is found in Image_A with 75%. The main reason for the lower precision rate is that the acquired image contains T-cell which have similar shape signature as DCs. Therefore, it is quite confusing to distinguish between both cells. Meanwhile, Image_B has the lowest recall rate with 66.7% as the region of one DC in the image is contaminated with bright halo region. Therefore, it is confusing to extract the cell region in the pre-processing method. Such issue can be improved later by enhancing the visualisation technique of PCM tool with proper way.
181
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182 Table1. Comparison result for CasDC and manual
Image
Image_D
Image_A
Image_B
Image_C
TP
3
2
4
3
TN
64
197
195
62
FP
1
0
1
0
FN
1
1
0
0
Precision (%)
75
100
80
100
Recall (%)
75
66.7
100
100
Accuracy (%)
97.1
99.5
99.5
100
Time (secs)
1.982
2.058
2.593
2.281
Over 145 images, the CasDC system achieves 64.9%, 71.7% and 87.8% of precision, recall and accuracy respectively as shown in Table 2. The system gives the lowest performance in term of precision and recall with 64.9% and 71.7% respectively. The main reason for the lower recall and accuracy is that the false positives are mainly from the clumping debris and other cells in the Data-set 2 as shown in Fig. 7. Moreover, the counting, particularly in Data-set 2 contains multiple redundancies on already counted cells. The system miscounts DCs on the areas with halo region and lower colour intensity as illustrated in Fig. 8(a) and (b). The pre-processing is unable to segment the DCs as the region is overlapped with T-cell and halo region. Therefore, the halo region is removed which cause the DCs contour to be separated. Table 2. Overall system performance for 2 Data-set Image
Precision (%)
Recall (%)
Accuracy (%)
Data-set 1
83.8
94.2
99.5
Data-set 2
62.4
68.7
86.4 T-cells with halo region
Multiple counting on DCs
Overlapping DCs with other cells
Fig. 7. Overlapping cells
Fig. 8(a). Pre-processing
Fig. 8(b). DCs counting
Based on the classifier results, the image quality can be enhanced in image acquisition stage, specifically in several parameter settings such lighting condition and lens. The technique used in the specimen preparation should be improved to reduce clumping cells such as debris which can deteriorate the counting process. Towards the end, to advance the system precision and recall, the problem of false positives could be alleviated using better preprocessing techniques and improve the number of templates used in the DCs library. Therefore, the system can be accurately identifying DCs with various parameters that control the image.
182
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
4. Conclusion In this paper, an image based computer aided system for cancer immunotherapy is developed and proposed for enhancing the vaccine preparation with Dendritic Cells (DCs) immunotherapy. The system consists of an image processing algorithm, which is able to discriminate and recognise DCs based on the cell shape signatures with low computational time. Moreover, the image data analysed by CasDC can be accessed by pathologists, doctors, and other clinical practitioners for proper cancer treatment. Next, CasDC is a system which requires a minimal configuration and almost low maintenance costs, unlike the Flow Cytometry machine used in the laboratory. The system achieves overall precision, recall, accuracy of 64.9%, 71.7%, and 87.8% respectively. In future, the image processing algorithm can be updated by advancing the method of feature extraction to overcome the recognition error rate that can limit the performance. The uses of advanced artificial intelligence, such as a deep neural network with numerous features of DCs should be investigated. Following that, the system feasibility can be enhanced by integrating with the recently launched concept of cloud computing [12]. Thus, the usability of CasDC can be widely applied in any working area. Another future extension for CasDC system can integrate biometrics of the individual patient and allow customized treatment for cancer, such as proper dosage delivered to the cancerous site. Acknowledgements This study is supported by Cancer Research Malaysia (CRM), Universiti Tun Hussein Onn Malaysia (UTHM) and Fundamental Research Grant (FRGS) under the Ministry of Higher Education (Vot 1583). References 1. Sasada, A., Takagi, M., Tabata, S., Abe, M. and Abe, H. A patient with stage IV gastric cancer who acquired complete remission after undergoing multi-peptide dendritic cell immunotherapy in combination with standard therapies. Personalized Medicine Universe 2015; 4: 7072. 2. Raïch-Regué, D., Glancy, M. and Thomson, A.W. Regulatory dendritic cell therapy: from rodents to clinical application. Immunology Letters 2014; 161(2): 216-221. 3. Conejo-Garcia, J.R., Rutkowski, M.R. and Cubillos-Ruiz, J.R. State-of-the-art of regulatory dendritic cells in cancer. Pharmacology & therapeutics 2016; 164:97-104. 4. Tan, Y.F., Leong, C.F. and Cheong, S.K. Observation of dendritic cell morphology under light, phase-contrast or confocal laser scanning microscopy. The Malaysian Journal of Pathology 2010; 32(2): 97-102. 5. Suberi, A.A.M., Zakaria, W.N.W., Tomari, R. and Lau, M.X., 2016, July. Dendritic cell recognition using template matching based on onedimensional (1D) Fourier descriptors (FD). In First International Workshop on Pattern Recognition. SPIE. 2016. pp. 100110K. 6. Hadjiiski, L., Sahiner, B. and Chan, H.P. Advances in CAD for diagnosis of breast cancer. Current opinion in obstetrics & gynecology 2006; 18(1): 64. 7. Dromain, C., Boyer, B., Ferre, R., Canale, S., Delaloge, S. and Balleyguier, C. Computed-aided diagnosis (CAD) in the detection of breast cancer. European journal of radiology 2013; 82(3): 417-423. 8. Shah, S., Dhameliya, V. and Roy, A.K. ImPatho-an image processing based pathological decision support system for disease identification and a novel tool for overall health governance. In Humanitarian Technology Conference (R10-HTC). IEEE. 2014. pp. 64-69. 9. Jaccard, N., Griffin, L.D., Keser, A., Macown, R.J., Super, A., Veraitch, F.S. and Szita, N. Automated method for the rapid and precise estimation of adherent cell culture characteristics from phase contrast microscopy images. Biotechnology and Bioengineering 2014; 111(3): 504-517. 10. Suberi, A.A.M., Zakaria, W.N.W., Tomari, R. and Lim, K.P. Optimization of overlapping dendritic cell segmentation in phase contrast micorscopy images. In 2016IEEE EMBS Conference on Biomedical Engineering and Sciences, IEEE. (Unpublished). 11. Tomari, R., Zakaria, W.N.W., Jamil, M.M.A., Nor, F.M. and Fuad, N.F.N. Computer aided system for red blood cell classification in blood smear image. Procedia Computer Science 2014; 42: 206-213. 12. Lee, H. and Chen, Y.P.P. Image based computer aided diagnosis system for cancer detection. Expert Systems with Applications 2015; 42(12): 5356-5365.
ScienceDirect Procedia Computer Science 105 (2017) 177 – 182
2016 IEEE International Symposium on Robotics and Intelligent Sensors, IRIS 2016, 17-20 December 2016, Tokyo, Japan
Dendritic Cell Recognition in Computer Aided System for Cancer Immunotherapy Anis Azwani Muhd Suberiª, Wan Nurshazwani Wan Zakaria*ª, Razali Tomariª ªAdvance Mechatronic Research Group (ADMIRE), Faculty of Electrical and ElectronicEngineering, Universiti Tun Hussein Onn Malaysia, Parit Raja Batu Pahat 86400, Malaysia
Abstract Immunotherapy is an entirely advanced class of cancer treatment which has been highly active and exciting field in clinical therapeutics. In numerous procedures, cancer immunotherapy demands a laborious practice to recognise and count Dendritic Cells (DCs) in the vaccine preparation. Conventionally, the laser-based technology that provides a rapid analysis such as Flow Cytometry can affect the DCs viability as the staining procedure is involved. Another highly promising method which is Phase Contrast Microscopy (PCM) involves experienced pathologists to visually examine the respective microscopy images. In facts, PCM confronts complex issues regarding to imaging artifacts which can deteriorate the recognition process. As the DCs counting is crucial in cancer treatment procedures, this paper proposes a pioneering system called CasDC which implement an image processing algorithm to recognise and count DCs with a label-free method. The aim of developing this system is to establish a reliable and time saving-tool as a second reader in the clinical practice. In the meantime, the treatment procedure can be administered and therefore, improve the patient’s survival rate. Our proposed system has an enormous potential towards helping Cancer Research Institute in which the system offers rapid and high throughput cancer immunotherapy vaccine preparation and automated live cell investigation. © 2017 byby Elsevier B.V. This is an open access article under the CC BY-NC-ND license © 2016 The TheAuthors. Authors.Published Published Elsevier B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors (IRIS2016). 2016). Sensors(IRIS Keywords: Dendritic cells; Cancer immunotherapy; Image processing; Pattern recognition; Phase contrast microsocopy; Computer vision
*
Corresponding author. Tel.: +6-07-453-7000; fax: +6-07-453-6060. E-mail address: [email protected]
1877-0509 © 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors(IRIS 2016). doi:10.1016/j.procs.2017.01.201
178
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
1.Introduction Most of the established therapies such as surgery, chemotherapy and radiation are widely used to treat cancer patients. Surgery [1] is often considered as the first therapy option to remove the tumour. Nevertheless, there is a tendency of removing only part of the tumour. Following that, a combination of surgery with radiation or chemotherapy is typically used in patient to kill the cancer cells [1]. However, both radiation and chemotherapy can introduce non-favourable outcomes towards the patient, such as serious bleeding, lack of energy and experience depression. Recently, immunotherapy has been widely explored and introduced as an advanced approach to boost the immune system to fight cancer [1–3]. Immunotherapy employs and activates the Dendritic Cells (DCs) to identify the cancer tissue in the normal body cells. DCs have a key role for activation of T- and B-cell immunity due to their superior ability to function as antigen-presenting cells (APCs) [2,4]. They can be generated in vitro from Peripheral Blood Mononuclear Cells (PBMCs) [4]. In DCs vaccine cell preparation, the identification of DCs from PBMCs sample is crucial before the cell can be stimulated. Conventional Flow Cytometry provides an effective recognition of labelled DCs using fluorescent dyes that may cause phototoxic damage to the DCs. Recent advances in cellular imaging have facilitated investigation of the unstained living cell using Phase Contrast Microscopy (PCM). Even though PCM is a label-free imaging modality, the identification purpose becomes challenging as the image is constituted with a variation of imaging artifacts such as halo region, low contrast and overlapping DCs [5]. Such procedure is laborious, time-consuming and very dependent on the expert’s skill to identify and count DCs which can introduce human errors. Previous researches introduce Computer Aided Diagnosis (CAD) [6,7] which is an active area of research and development in medical imaging and diagnostics. A system called ImPatho [8], has been developed for disease identification. This system helps to provide a proper clinical guideline towards any disease, such as RBC disorder. Although most clinical applications are developed for cancer detection, it can be expected that the DCs counting in immunotherapy will be an important component of cancer treatment. In this paper, the challenges and complex issues faced by Cancer Research Institute are addressed. Our proposed system, CasDC (Computer Aided System for Dendritic Cells), is developed to identify and count DCs in the sample. The main goal of CasDC system is to assist pathologists and other clinical practitioners in initiating the vaccine preparation for cancer immunotherapy treatment. The details of the proposed framework are described in Section 2, system overview, followed by the results and discussion in Section 3. Finally, the conclusion is summarized in Section 4. 2.System Overview In this section, the CasDC workflow is presented in Fig. 1. It consists of two stages; a) Image acquisition and b) image processing in CasDC. The system is developed, trained and tested using Matlab R2015a platform and runs on 2.5GHz i3-312M processor. The image dataset is provided by Cancer Research Malaysia (CRM) which composed of debris (black dots), T-cells (round shape) and DCs (red circle) as shown in Fig. 2. The digital images of PBMCs are captured by Nikon DS-Fi2 camera which attached to Olympus CK40 microscope with 100x and 200x magnification. All the images have variations of light-variant environments. PBMCs sample
Microscope Image acquisition
Image normalization
Pre-processing
Classification
Feature Extraction
Segmentation
Blood smear
Fig. 1. CasDC workflow
Fig. 2 PCM image sample
179
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
2.1.Workflow of CasDC Image Analysis Fig. 3 demonstrates the workflow of CasDC in DCs recognition and counting. Fig. 3(a) and (b) show the sample images are prepared for processing and counting process. A cell library that composed of DCs and other cells (T-cell and debris) is provided where it acts as a template which will be compared to the testing image to detect the most approximate cell signatures. During the matching process, One Dimensional (1D) Fourier Descriptors computation is used to compare the template with the testing image, which can enable rapid and efficient template matching to identified, mark and count DCs as shown in Fig. 3(c) and (d). (a)
Start Insert
(b)
Templates and testing images
Pre-processing and Segmentation
(d) (c)
Template matching based on 1D Fourier Descriptors
DCs recognition, marking and counting
Sample image
Display count result
End
Fig. 3. Workflow of CasDC
2.2.Algorithm for Dendritic Cells (DCs) Detection The key element of CasDC system is an image processing algorithm for recognising DCs as shown in Fig. 4. As an initial step, image normalization is applied to each testing image to reduce the illumination effect. Based on the previous work , a combination of Local Contrast Threshold (LCT) with halo removal [9] is applied as the PCM images are contaminated with imaging artifacts. However, the image contains overlapping and clumping cells which can deteriorate the classification process. To overcome these problems, a hybrid of low and high sigma in Gaussian kernel filtering through logical operator AND with Local Adaptive Threshold (H-GLAT) [10] have been applied in which this method address the imaging artifacts issues in overlapping and clumping cells as shown in Fig. 4(b). Subsequently, the segmentation masks are post-processed by several morphology operations and Canny edge to segregate the entire cell shapes in the image as shown in Fig. 4(c) and (d) respectively. The Fourier Descriptors (FDs) have been applied for recognition purposes and this method involves One-Dimensional (1D) version to represent the contour in the form of shape signatures. Towards the end, the classification stage is performed by using template matching [5]. The identified DCs are marked with red colour as shown in Fig. 4(e).
a)
b)
c)
d)
e)
Fig. 4. a) Original image b) Pre-processing (H-GLAT) c) Morphological operators d) Canny edge and e) DCs detection
180
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
2.3.Graphical User Interface (GUI) Overview Graphical user interface (GUI) has been developed for CasDC to facilitate the pathologist to identify and count DC as shown in Fig. 5. This automated system provides user-friendly features to the user in which the sample images can be processed into two modes; 1) single and 2) batch. Both modes can be operated for 100x and 200x magnification. Once the mode is selected, the user needs to press ‘Start Analysis’ button which will load image, process and extract the cell shape. The system continuously generates Fourier Descriptors (FDs) through the entire image to find the approximate similarity between the DCs template and the sample image. For a single image, the user needs to save the result manually using ‘Save Result’ button. On the other hand, for batch file mode, the system will continuously scan, mark and count DCs in the entire file of images and automatically save the results once the process is completed as shown in Fig. 6.
Fig. 5. CasDC: DCs recognition and counting GUI
Fig. 6. Result of batch mode
3.Results and Discussion Two image dataset in the JPEG picture format are used for evaluating system performance. The first and second dataset are composed of 1877x1408 (n=10) and 2560x1920 (n=135) pixels images respectively. Both image dataset consists with variations of image colour, light environments and shapes of DCs. Table 1 show some of the comparison of DCs counting between CasDC and manual with processing time. The identified DCs of CasDC and manual are marked with red colour and black arrow respectively. Meanwhile, the average processing time to run the system for each image is 2.229 seconds. Next, the quantitative measurement is performed based on a) precision, b) recall and c) accuracy as shown in Equation 1 [11]. Precision refers to the exact amount of DCs and recall presents the exact number of DCs from the input image. Meanwhile, accuracy measures the overall system performance in classifying DCs. Pr ecision =
TP TP TP + TN ; Re call = ; Accuracy = TP + FP TP + FN TP + FP + TN + FN
(1)
The lowest performance in term of precision is found in Image_A with 75%. The main reason for the lower precision rate is that the acquired image contains T-cell which have similar shape signature as DCs. Therefore, it is quite confusing to distinguish between both cells. Meanwhile, Image_B has the lowest recall rate with 66.7% as the region of one DC in the image is contaminated with bright halo region. Therefore, it is confusing to extract the cell region in the pre-processing method. Such issue can be improved later by enhancing the visualisation technique of PCM tool with proper way.
181
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182 Table1. Comparison result for CasDC and manual
Image
Image_D
Image_A
Image_B
Image_C
TP
3
2
4
3
TN
64
197
195
62
FP
1
0
1
0
FN
1
1
0
0
Precision (%)
75
100
80
100
Recall (%)
75
66.7
100
100
Accuracy (%)
97.1
99.5
99.5
100
Time (secs)
1.982
2.058
2.593
2.281
Over 145 images, the CasDC system achieves 64.9%, 71.7% and 87.8% of precision, recall and accuracy respectively as shown in Table 2. The system gives the lowest performance in term of precision and recall with 64.9% and 71.7% respectively. The main reason for the lower recall and accuracy is that the false positives are mainly from the clumping debris and other cells in the Data-set 2 as shown in Fig. 7. Moreover, the counting, particularly in Data-set 2 contains multiple redundancies on already counted cells. The system miscounts DCs on the areas with halo region and lower colour intensity as illustrated in Fig. 8(a) and (b). The pre-processing is unable to segment the DCs as the region is overlapped with T-cell and halo region. Therefore, the halo region is removed which cause the DCs contour to be separated. Table 2. Overall system performance for 2 Data-set Image
Precision (%)
Recall (%)
Accuracy (%)
Data-set 1
83.8
94.2
99.5
Data-set 2
62.4
68.7
86.4 T-cells with halo region
Multiple counting on DCs
Overlapping DCs with other cells
Fig. 7. Overlapping cells
Fig. 8(a). Pre-processing
Fig. 8(b). DCs counting
Based on the classifier results, the image quality can be enhanced in image acquisition stage, specifically in several parameter settings such lighting condition and lens. The technique used in the specimen preparation should be improved to reduce clumping cells such as debris which can deteriorate the counting process. Towards the end, to advance the system precision and recall, the problem of false positives could be alleviated using better preprocessing techniques and improve the number of templates used in the DCs library. Therefore, the system can be accurately identifying DCs with various parameters that control the image.
182
Anis Azwani Muhd Suberi et al. / Procedia Computer Science 105 (2017) 177 – 182
4. Conclusion In this paper, an image based computer aided system for cancer immunotherapy is developed and proposed for enhancing the vaccine preparation with Dendritic Cells (DCs) immunotherapy. The system consists of an image processing algorithm, which is able to discriminate and recognise DCs based on the cell shape signatures with low computational time. Moreover, the image data analysed by CasDC can be accessed by pathologists, doctors, and other clinical practitioners for proper cancer treatment. Next, CasDC is a system which requires a minimal configuration and almost low maintenance costs, unlike the Flow Cytometry machine used in the laboratory. The system achieves overall precision, recall, accuracy of 64.9%, 71.7%, and 87.8% respectively. In future, the image processing algorithm can be updated by advancing the method of feature extraction to overcome the recognition error rate that can limit the performance. The uses of advanced artificial intelligence, such as a deep neural network with numerous features of DCs should be investigated. Following that, the system feasibility can be enhanced by integrating with the recently launched concept of cloud computing [12]. Thus, the usability of CasDC can be widely applied in any working area. Another future extension for CasDC system can integrate biometrics of the individual patient and allow customized treatment for cancer, such as proper dosage delivered to the cancerous site. Acknowledgements This study is supported by Cancer Research Malaysia (CRM), Universiti Tun Hussein Onn Malaysia (UTHM) and Fundamental Research Grant (FRGS) under the Ministry of Higher Education (Vot 1583). References 1. Sasada, A., Takagi, M., Tabata, S., Abe, M. and Abe, H. A patient with stage IV gastric cancer who acquired complete remission after undergoing multi-peptide dendritic cell immunotherapy in combination with standard therapies. Personalized Medicine Universe 2015; 4: 7072. 2. Raïch-Regué, D., Glancy, M. and Thomson, A.W. Regulatory dendritic cell therapy: from rodents to clinical application. Immunology Letters 2014; 161(2): 216-221. 3. Conejo-Garcia, J.R., Rutkowski, M.R. and Cubillos-Ruiz, J.R. State-of-the-art of regulatory dendritic cells in cancer. Pharmacology & therapeutics 2016; 164:97-104. 4. Tan, Y.F., Leong, C.F. and Cheong, S.K. Observation of dendritic cell morphology under light, phase-contrast or confocal laser scanning microscopy. The Malaysian Journal of Pathology 2010; 32(2): 97-102. 5. Suberi, A.A.M., Zakaria, W.N.W., Tomari, R. and Lau, M.X., 2016, July. Dendritic cell recognition using template matching based on onedimensional (1D) Fourier descriptors (FD). In First International Workshop on Pattern Recognition. SPIE. 2016. pp. 100110K. 6. Hadjiiski, L., Sahiner, B. and Chan, H.P. Advances in CAD for diagnosis of breast cancer. Current opinion in obstetrics & gynecology 2006; 18(1): 64. 7. Dromain, C., Boyer, B., Ferre, R., Canale, S., Delaloge, S. and Balleyguier, C. Computed-aided diagnosis (CAD) in the detection of breast cancer. European journal of radiology 2013; 82(3): 417-423. 8. Shah, S., Dhameliya, V. and Roy, A.K. ImPatho-an image processing based pathological decision support system for disease identification and a novel tool for overall health governance. In Humanitarian Technology Conference (R10-HTC). IEEE. 2014. pp. 64-69. 9. Jaccard, N., Griffin, L.D., Keser, A., Macown, R.J., Super, A., Veraitch, F.S. and Szita, N. Automated method for the rapid and precise estimation of adherent cell culture characteristics from phase contrast microscopy images. Biotechnology and Bioengineering 2014; 111(3): 504-517. 10. Suberi, A.A.M., Zakaria, W.N.W., Tomari, R. and Lim, K.P. Optimization of overlapping dendritic cell segmentation in phase contrast micorscopy images. In 2016IEEE EMBS Conference on Biomedical Engineering and Sciences, IEEE. (Unpublished). 11. Tomari, R., Zakaria, W.N.W., Jamil, M.M.A., Nor, F.M. and Fuad, N.F.N. Computer aided system for red blood cell classification in blood smear image. Procedia Computer Science 2014; 42: 206-213. 12. Lee, H. and Chen, Y.P.P. Image based computer aided diagnosis system for cancer detection. Expert Systems with Applications 2015; 42(12): 5356-5365.