A computer-aided diagnosis system for breast pathology: a deep learning approach with model interpretability from pathological perspective

• Published in: Service oriented computing and applications 2024-06, Vol.18 (2), p.183-193
• Main Authors: Hsu, Wei-Wen, Wu, Yongfang, Chen, Chung-Hao, Hao, Chang, Hou, Yu-Ling, Gao, Xiang, Shao, Yun, Zhang, Xueli, He, Tao, Tai, Yanhong
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial neural networks; Breast cancer; Classification; Computer Appl. in Administrative Data Processing; Computer Science; Computer Systems Organization and Communication Networks; Diagnosis; e-Commerce/e-business; Explainable artificial intelligence; IT in Business; Lesions; Machine learning; Management of Computing and Information Systems; Performance evaluation; Software Engineering/Programming and Operating Systems; Special Issue Paper
• ISSN: 1863-2386; 1863-2394
• DOI: 10.1007/s11761-023-00378-4

In this study, a computer-aided diagnosis (CAD) system for whole-slide images (WSIs) with breast cancer was proposed to achieve lesion detection and classification. Different from the previously proposed works that performed segmentation directly and applied the segmentation metrics for performance evaluation, here we computed the accuracy in patch-level, region-level, and slide-level to access our proposed framework to meet the clinical requirements, instead of simply regarding it as an AI segmentation task. To achieve that for WSIs, lesion detection was conducted at high magnification (× 40) first for region proposal by patching using the modified AlexNet model; and then, several patches were sampled from the detected regions at lower magnification (× 10) to observe cells’ patterns for lesion classification by the model of ResNet50. With such fashion, the deep features being distinguishing in lesion classification from the convolutional neural networks (CNN) were further analyzed in this work to provide comprehensive interpretability for the proposed CAD system based on pathological knowledge. For our experiments, a total of 186 slides of WSIs were collected and classified into three categories: Non-Carcinoma, Ductal Carcinoma in Situ (DCIS), and Invasive Ductal Carcinoma (IDC). The slide-level accuracy rate for the three-category classification reached 90.8% (99/109) through a fivefold cross-validation and achieved 94.8% (73/77) on the testing set. In addition, the learned features reflect the clinical insights in lesion classification for explainable AI, which enhanced the reliability and validity of our proposed system.