BreastCare AI – Deep Learning Breast Cancer Detection

Key Highlights

• Built a full-stack AI web application for breast cancer detection using a fine-tuned Swin Transformer trained on the CBIS-DDSM mammography dataset (~30,000 images), achieving 87.5% accuracy and 0.884 ROC-AUC.
• Implemented a dual-view ensemble analysis system (CC + MLO views) mimicking clinical radiologist workflow, where probability averaging across views reduces false positives and improves diagnostic confidence.
• Developed a FastAPI backend with optimized model inference pipeline, LRU caching for repeated scans, and asynchronous file handling designed for high-throughput clinical usage.
• Built a React frontend featuring drag-and-drop DICOM/JPEG upload, real-time prediction results with probability heatmaps, and automated PDF report generation with a clinical findings summary.
• Applied transfer learning with an ImageNet-pretrained Swin Transformer, class weighting to address dataset imbalance, and augmentation techniques (flips, rotations, contrast jitter) for robust generalization.
• Hosted the model on HuggingFace Hub for version-controlled model management; containerized the full stack with Docker Compose for reproducible deployment in clinical environments.