Anatomy-aware routing for robust medical image segmentation: a failure-detectable alternative to mixed models

Official URL: https://dx.doi.org/10.1109/SIU71813.2026.11636837

Heterogeneous medical segmentation pipelines that use a single shared model lose organ-specific feature learning and degrade silently under distribution shift. We propose an anatomy-aware framework where an upstream classifier routes inputs to dedicated specialist networks, combined with a composite post-hoc failure detector based on MC-Dropout uncertainty, domain-shift detection, and plausibility checking. Using chest X-ray pneumothorax and brain MRI datasets, we show anatomy-aware routing yields higher segmentation accuracy (+8.6-14.4 Dice points) compared to a mixed-organ baseline. Our failure detector achieves flag rates of 1.000 on misrouted pneumothorax cases and 0.963 on brain cases. In contrast, under identical noise, the mixed model produces flag rates of only 0.173 and 0.156. Anatomy-aware pipelines offer superior accuracy and the ability to fail loudly, allowing safer clinical deployment through explicit abstention.