Back to projects
Multi-Modal Breast Cancer Detection
Deep learning framework for breast cancer detection across ultrasound, mammography and histology images, with attention- and gated-fusion across modalities.
Overview
A framework supporting single-modality analysis and combined multi-modal approaches for breast cancer detection.
My role
Researcher
Problem
Detecting breast cancer across several imaging modalities — ultrasound, mammography and histology — within one framework.
Approach
Supports 10+ backbones and three fusion methods (attention, gated, concatenation). Training uses mixed precision, AdamW with cosine scheduling, early stopping and modality-specific augmentation on a 70/15/15 split; evaluation includes bootstrap confidence intervals and calibration metrics.
Results
- Accuracy: ultrasound 94.2% (EfficientNet-B4), mammography 91.8% (EfficientNet-B0), histology 93.5% (Swin-T), multi-modal attention fusion 96.1%.
Multi-modal accuracy
96.1%
Multi-modal AUC
0.98
Tech stack
PyTorchEfficientNetSwin-TMulti-modal fusionMedical imaging