CascadePLS-ViT: Cascade with Patch-Level Self-Supervised Vision Transformers for Breast Cancer Classification in Mammography

This paper introduces CascadePLS-ViT, a novel cascade framework for classifying Breast Cancer (BC) into three BI-RADS categories using mammogram images. Our method leverages self-supervised learning (SSL) and Vision Transformers (ViTs) to analyze image patches, providing a comprehensive understanding of mammographic features crucial for accurate BI-RADS categorization. The first stage of the cascade differentiates between non-cancerous and potentially cancerous mammograms using a unique self-supervised task, SelfPatch. This task enhances patch-level representation learning without the need for annotated datasets by enforcing invariance against each patch and its neighboring patches, treating similar neighboring patches as positive samples. The second stage of the cascade refines the classification by distinguishing between the Scattered Fibroglandular class and the Heterogeneously, Extremely Dense class. This detailed classification enables precise and nuanced breast cancer grading. Our integrated approach combines the granularity of patch-level analysis with the robustness of cascade classification, ensuring high Recall and Accuracy across BI-RADS categories. Rigorous testing on 4368 patients across three BI-RADS classes yielded a system Recall of 85.01% and an F1 score of 84.90%.