Annotation-efficient deep learning for breast cancer whole-slide image classification using Tumor Infiltrating Lymphocytes and slide-level labels

Abstract Tumor-Infiltrating Lymphocytes (TILs) play a crucial role in the immune response against cancer cells. The existing deep learning methods for TIL analysis in whole-slide images (WSIs) require extensive annotations at the sub-image (patch) level, which often require time-consuming and labor-intensive specialist input. To address this, we propose a first-of-its-kind framework named annotation-efficient segmentation and attention-based classifier (ANSAC). ANSAC requires only slide-level labels to classify WSIs as having high vs. low TIL scores, with the binary classes divided by an expert-defined threshold. ANSAC automatically segments tumor and stroma regions relevant to TIL assessment, eliminating the need for extensive manual annotations. Additionally, it uses an attention model to generate an attention map that highlights the most pertinent regions in the WSI for the classification task. We evaluate ANSAC on four breast cancer WSI datasets and show significant improvements over three baseline methods in identifying TIL-relevant regions and up to 8% classification improvement on a held-out test dataset. Additionally, we propose a modification to the pre-processing step of a well-known method, enhancing its performance up to 6%.