Standard 3D U-Net beats handcrafted shape priors in cardiac CT segmentation
New study finds simple U-Net outperforms explicit anatomical priors on two heart datasets.
A team led by Michael Hudler from the Medical University of Graz (published at AIRoV 2026) set out to test whether explicit anatomical shape priors could improve whole-heart multi-compartment CT segmentation—a clinically critical task. They implemented shape-aware losses and spatial label distribution heatmap-guided U-Net variants, drawing statistics directly from training data, and benchmarked these against a standard 3D U-Net on the MM-WHS CT and WHS++ datasets.
Surprisingly, across all experiments, the standard 3D U-Net remained a very strong baseline. The handcrafted shape priors produced at best marginal and inconsistent changes and often degraded segmentation performance. The authors conclude that the baseline already captures substantial implicit anatomical regularities, and that future gains will likely require more expressive learned priors—for example, from generative models or transformers—rather than simple handcrafted constraints.
- Standard 3D U-Net outperformed or matched all handcrafted shape priors on MM-WHS CT and WHS++ datasets.
- Shape-aware losses and spatial label distribution heatmap guidance yielded only marginal, inconsistent improvements and sometimes hurt accuracy.
- Authors recommend focusing on more expressive learned priors (e.g., from generative models) for future cardiac segmentation advances.
Why It Matters
Challenges the assumption that explicit anatomical constraints improve medical AI; future work should shift to learned priors.