Building and Comparing Segmentation Strategies for Coronary Artery Calcium CAC

CAC segmentation is harder than it looks. Calcium deposits are tiny, sparse, and look identical to nearby bone a small miss doesn't just drop your Dice score, it flips a patient's risk category entirely. I already ran a test: a lightweight 2D U-Net hitting 0.9416 median Dice, 63x faster than TotalSegmentator. Strong baseline, but high variance on edge cases showed me exactly where it breaks. So instead of optimizing one model, I'm building a reproducible pipeline to compare five architectures 2D U-Net, 2.5D U-Net, Attention U-Net, TransUNet/SwinUNet, and nnU-Net under identical conditions. Same data, same loss, same optimizer. Only the architecture changes. That way the results actually mean something. Beyond Dice scores, I'll evaluate Agatston score accuracy, risk category agreement, and anatomical localization of plaques to specific coronary arteries. The goal is a framework the PREDICT1 team can actually use not just a leaderboard.