Enhancing Breast Cancer Classification with Stacked Convolutional Neural Networks for Whole-Slide Image Analysis

This study explores an advanced approach to automatic breast cancer characterization by leveraging convolutional neural networks (CNNs) trained on 224×224 pixel patches from whole-slide images (WSIs). I compared two architectures, VGG-16 and WRN-4-2, across two classification problems: benign versus cancer, and benign, DCIS, versus IDC. Results indicated that WRN-4-2 slightly outperformed VGG-16 on the two-class problem but underperformed on the three-class problem. Stacking CNNs with increased input sizes (512×512, 768×768, and 1024×1024 pixels) significantly improved accuracy, with the 1024×1024 network achieving the highest performance but at a higher computational cost. Dense prediction maps were generated from these stacked networks, enabling effective whole-slide image classification with approximately 90% accuracy for detecting cancer. However, performance for distinguishing between benign, DCIS, and IDC lesions was lower at 76.6%. Future improvements could involve stain standardization, enhanced resolution of prediction maps using architectures like U-net, and comparative studies of multi-resolution approaches to further refine cancer detection and classification accuracy.