Muhammad Mohsin Zafar1, Zunaira Rauf2, Anabia Sohail2, Abdul Rehman Khan3, Muhammad Obaidullah3, Saddam Hussain Khan2, Yeon Soo Lee4, Asifullah Khan5. 1. Pattern Recognition Lab, Department of Computer & Information Sciences, Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan; Faculty of Computer Science and Engineering, Ghulam Ishaq Khan Institute of Engineering Sciences and Technology, Topi 23640, District Swabi, Khyber Pakhtunkhwa, Pakistan. 2. Pattern Recognition Lab, Department of Computer & Information Sciences, Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan; PIEAS Artificial Intelligence Center (PAIC), Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan. 3. Pattern Recognition Lab, Department of Computer & Information Sciences, Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan; Center for Mathematical Sciences, Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan. 4. Deparment of Biomedical Engineering, College of Medical Sciences, Catholic University of Daegu, South Korea. Electronic address: yeonsoolee@cu.ac.kr. 5. Pattern Recognition Lab, Department of Computer & Information Sciences, Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan; PIEAS Artificial Intelligence Center (PAIC), Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan; Deparment of Biomedical Engineering, College of Medical Sciences, Catholic University of Daegu, South Korea; Center for Mathematical Sciences, Pakistan Institute of Engineering & Applied Sciences, Nilore, Islamabad 45650, Pakistan. Electronic address: asif@pieas.edu.pk.
Abstract
BACKGROUND: Immuno-score, a prognostic measure for cancer, employed in determining tumor grade and type, is generated by counting the number of Tumour-Infiltrating Lymphocytes (TILs) in CD3 and CD8 stained histopathological tissue samples. Significant stain variations and heterogeneity in lymphocytes' spatial distribution and density make automated counting of TILs' a challenging task. METHODS: This work addresses the aforementioned challenges by developing a pipeline "Two-Phase Deep Convolutional Neural Network based Lymphocyte Counter (TDC-LC)" to detect lymphocytes in CD3 and CD8 stained histology images. The proposed pipeline sequentially works by removing hard negative examples (artifacts) in the first phase using a custom CNN "LSATM-Net" that exploits the idea of a split, asymmetric transform, and merge. Whereas, in the second phase, instance segmentation is performed to detect and generate a lymphocyte count against the remaining samples. Furthermore, the effectiveness of the proposed pipeline is measured by comparing it with the state-of-the-art single- and two-stage detectors. The inference code is available at GitHub Repository https://github.com/m-mohsin-zafar/tdc-lc. RESULTS: The empirical evaluation on samples from LYSTO dataset shows that the proposed LSTAM-Net can learn variations in the images and precisely remove the hard negative stain artifacts with an F-score of 0.74. The detection analysis shows that the proposed TDC-LC outperforms the existing models in identifying and counting lymphocytes with high Recall (0.87) and F-score (0.89). Moreover, the commendable performance of the proposed TDC-LC in different organs suggests a good generalization. CONCLUSION: The promising performance of the proposed pipeline suggests that it can serve as an automated system for detecting and counting lymphocytes from patches of tissue samples thereby reducing the burden on pathologists.
BACKGROUND: Immuno-score, a prognostic measure for cancer, employed in determining tumor grade and type, is generated by counting the number of Tumour-Infiltrating Lymphocytes (TILs) in CD3 and CD8 stained histopathological tissue samples. Significant stain variations and heterogeneity in lymphocytes' spatial distribution and density make automated counting of TILs' a challenging task. METHODS: This work addresses the aforementioned challenges by developing a pipeline "Two-Phase Deep Convolutional Neural Network based Lymphocyte Counter (TDC-LC)" to detect lymphocytes in CD3 and CD8 stained histology images. The proposed pipeline sequentially works by removing hard negative examples (artifacts) in the first phase using a custom CNN "LSATM-Net" that exploits the idea of a split, asymmetric transform, and merge. Whereas, in the second phase, instance segmentation is performed to detect and generate a lymphocyte count against the remaining samples. Furthermore, the effectiveness of the proposed pipeline is measured by comparing it with the state-of-the-art single- and two-stage detectors. The inference code is available at GitHub Repository https://github.com/m-mohsin-zafar/tdc-lc. RESULTS: The empirical evaluation on samples from LYSTO dataset shows that the proposed LSTAM-Net can learn variations in the images and precisely remove the hard negative stain artifacts with an F-score of 0.74. The detection analysis shows that the proposed TDC-LC outperforms the existing models in identifying and counting lymphocytes with high Recall (0.87) and F-score (0.89). Moreover, the commendable performance of the proposed TDC-LC in different organs suggests a good generalization. CONCLUSION: The promising performance of the proposed pipeline suggests that it can serve as an automated system for detecting and counting lymphocytes from patches of tissue samples thereby reducing the burden on pathologists.