TRACE-Seg3D: Counterfactual Context Auditing For Robust 3D Glioma Segmentation Under Institutional Shift

Medical image segmentation models often perform well on standard benchmarks but fail unpredictably when applied to data from different scanners or institutions. Conventional metrics like Dice and HD95 measure overlap but fail to explain why a model fails or whether a prediction is driven by actual tumor evidence versus spurious imaging context. This paper asks: how can we build 3D segmentation models that are not only accurate but also auditable, providing evidence of their reliability under distribution shift?

TRACE-Seg3D introduces a causal framework that treats the segmentation task as a combination of disease evidence and imaging context. The model uses a proxy-anchored encoder to factorize features into disease-guided and context-guided latents. To audit performance, the authors implement Counterfactual Context Transport (CCT), which holds the disease evidence constant while systematically varying the imaging context using a support bank of other cases. This allows the model to generate a consensus prediction and a voxel-wise instability map, highlighting regions where the model's output is sensitive to context changes. Additionally, the framework enforces an anatomical structural prior (ET ⊆ TC ⊆ WT) to ensure that the predicted subregions are biologically plausible.

Experiments on BraTS and UTSW glioma datasets demonstrate that TRACE-Seg3D achieves competitive performance in both in-distribution and cross-domain settings. Crucially, the framework successfully exposes context-sensitive failure modes that are typically missed by standard overlap metrics. By providing an instability map and a stability-gated mask, the model allows clinicians to identify which parts of a segmentation are robust and which are potentially artifacts of the specific scanner or protocol used, offering a more transparent path toward clinical deployment.