Hybrid 3D CNN with Deformable Attention and FNO for CAC Segmentation

This project develops a high-precision, clinically viable pipeline for segmenting Coronary Artery Calcium (CAC) from non contrast cardiac CT scans using the Stanford COCA dataset. The core innovation is a Hybrid 3D U Net architecture that balances the local inductive bias of CNNs with the global context of Transformers, specifically optimized for the extreme sparsity of calcified lesions. The architecture integrates 3D Deformable Attention to bypass O(N^2) scaling and CoordConv for anatomical grounding, effectively suppressing false positives from osseous structures like ribs. A primary research contribution is the experimental Fourier Neural Operator (FNO) bottleneck, which tests the hypothesis that calcified plaques can be isolated as high-frequency signals in the spectral domain. To ensure clinical utility, the model is trained using Scoring-Driven Consistency (SDC) a differentiable volume score loss that aligns optimization with the Agatston metric. Deliverables include a MONAI based preprocessing pipeline, a rigorous benchmark against an nnUNet gold standard, and a fully automated Agatston scoring engine for risk categorization (0, 1–99, 100–399, ≥400). This work aims to bridge the gap between architectural novelty and real world diagnostic accuracy for cardiovascular risk stratification.