BRAIN (Breast Retrieval-Augmented Intelligence Network)

Designed an explainable AI system for breast cancer diagnosis utilizing a Swin Transformer Base (Swin-B) architecture.

Technical Innovation

Designed a custom Swin-B Transformer model equipped with a Dual-Head Loss system to analyze mammography data. Key technical features include:

Dual-Head Loss: Framework combining Cross-Entropy and Triplet Loss with Batch Hard Mining to optimize feature extraction.
RAG-based Verification: Engineered a Retrieval-Augmented Verification module using Facebook's FAISS library to cross-check predictions against a 24,576-image knowledge base (IMSMD).
Lesion Visualization: Integrated Grad-CAM++ and Attention Rollout to provide visual heatmaps of pathological features like microcalcifications and spiculated margins.

Achieved 100% classification accuracy and an AUC-ROC of 1.00 on a held-out test set of 4,916 multi-institutional mammograms.
Delivered a clinical decision support system that surfaces transparent visualizations of lesion boundaries.
Mitigated AI hallucinations by verifying predictions against historical cases using the FAISS similarity search.