Deep Learning-Based Breast Cancer Subtype Classification from Whole-Slide Images: Leveraging the BRACS Dataset

Breast cancer is one of the global leading causes of death in women. Accurate diagnosis and effective classification of breast cancer subtypes are critical elements in guiding the treatment and care of patients to improve survival rates. For subtype classification, whole slide imaging (WSI) is becoming increasingly widespread as it allows high-resolution digital capture of tissue samples and facilitates more efficient storage, access, and analysis. However, due to the complexity and variability of breast tissue samples, as well as the inherent subjectivity of human interpretation, there is a need to develop automated and reliable approaches. In our work, we have addressed the challenge of breast cancer type and subtype classification using advanced Deep Learning techniques based on Convolutional Neural Networks (CNN). The central problem concerns the correct identification and categorization of distinct breast tissue lesions from images of interest regions of hematoxylin and eosin (H&E) stained slides extracted from the BReast Carcinoma Subtyping Data Set (BRACS). Advanced techniques such as fine-tuning with pre-trained CNNs, data augmentation, and dropout are employed to enhance the algorithm’s generalization and prevent overfitting. Our experimental results demonstrate outstanding accuracy in breast cancer subtype classification, achieving accurate classifications in 3 and 7 classes at the Region of interest (RoI) level. Our study aims to advance research in the field of AI applied to breast pathology, providing a valuable tool for future research and clinical applications, to ultimately benefit the treatment and care of breast cancer patients.