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

Technical Summary: TRACE-Seg3D

Problem Statement
Medical image segmentation models often achieve high benchmark performance (e.g., Dice scores) yet fail silently when deployed on data from different scanners, protocols, or institutions. These "context shifts" alter image appearance without changing the underlying lesion, causing models to exploit nuisance cues rather than genuine disease evidence. Conventional overlap metrics (Dice, HD95) fail to distinguish between lesion-grounded predictions and those driven by spurious context, nor do they reveal whether predicted tumor subregions (Whole Tumor, Tumor Core, Enhancing Tumor) obey necessary anatomical hierarchies. This lack of transparency hinders trust in clinical deployment under distribution shift.

Methodology: TRACE-Seg3D
The authors propose TRACE-Seg3D, a counterfactual context auditing framework designed to produce robust 3D glioma segmentations alongside case-level evidence regarding stability and anatomical plausibility. The framework operates through three aligned mechanisms:

1. Proxy-Anchored Disease/Context Factorization (M1):
   The model employs a 3D encoder-decoder (MedNeXt-style) that extracts hierarchical features. A bottleneck feature is globally pooled to generate two latent vectors: a disease latent ($z_d$) and a context latent ($z_c$).

   • Separation: $z_d$ is encouraged to encode lesion evidence, while $z_c$ encodes acquisition and institutional variations.
   • Supervision: Since true causal variables are unobserved, the model uses "weak proxy anchors." $z_d$ is supervised by disease proxies (e.g., lesion volume, burden), and $z_c$ by context proxies (e.g., site metadata, intensity statistics, or pseudo-proxies derived from image geometry).
   • Regularization: Adversarial objectives and orthogonality constraints prevent leakage between the two factors, ensuring $z_d$ does not encode context and vice versa. These latents modulate the decoder features via FiLM layers.

2. Counterfactual Context Transport (CCT) (M2):
   To audit prediction stability, the framework performs latent interventions.

   • Process: For a target case, the disease latent ($z_d$) is held fixed, while the context latent ($z_c$) is replaced with samples drawn from a "support bank" of context latents from other training cases.
   • Output: The model generates multiple predictions under these counterfactual contexts. It computes a consensus prediction ($p^{cct}$) and a voxel-wise instability map ($u^{cct}$), which quantifies the variance of predictions across the transported contexts.
   • Stability Gating: A stability-gated mask ($p^{gate}$) is produced by suppressing regions where the instability exceeds a threshold, flagging predictions that are highly sensitive to context changes.

3. Anatomy-Aware Structural Prior (M3):
   Recognizing that Enhancing Tumor (ET) is small and prone to false positives under domain shift, the framework enforces the anatomical hierarchy $ET \subseteq TC \subseteq WT$.

   • Mechanism: A structural operator ($\Pi_{anat}$) removes isolated ET components that lack sufficient support from the Tumor Core (TC) or Whole Tumor (WT) regions.
   • Audit: The system reports the volume of removed ET components and the preservation of TC/WT regions, providing a measure of anatomical plausibility.

Key Contributions

• Causal Perspective: The paper formalizes 3D glioma segmentation under distribution shift by separating disease evidence, imaging context, and anatomical segmentation, treating the latter as the target of auditability.
• Auditable Framework: TRACE-Seg3D is the first to combine proxy-anchored factorization with Counterfactual Context Transport to produce not just a segmentation mask, but explicit evidence of context sensitivity and structural consistency.
• Anatomical Constraints: It integrates a structural prior that enforces the nested tumor hierarchy, specifically controlling false positives in the clinically critical Enhancing Tumor (ET) region.

Experimental Results
The framework was evaluated on the BraTS 2020 and UTSW-Glioma datasets, testing both in-distribution (ID) and cross-domain (Out-of-Distribution, OOD) performance.

• Performance: TRACE-Seg3D achieved state-of-the-art results in both ID and OOD settings. For example, in the BraTS $\to$ UTSW transfer, it achieved a Dice score of 0.591 (vs. 0.567 for the best baseline) and an HD95 of 4.558. In the UTSW $\to$ BraTS transfer, it reached 0.792 Dice.
• Ablation: Removing components (factorization, CCT, or structural prior) consistently degraded OOD performance and audit quality, confirming that the gains arise from the integrated mechanism rather than a single post-processing step.
• Small Structure Analysis: The framework significantly improved ET precision and reduced false positives (up to 52.6% reduction in OOD settings) without compromising the segmentation of larger tumor regions (WT/TC).
• Audit Utility: The instability maps successfully highlighted regions where predictions were context-sensitive (e.g., boundaries or small foci), providing a visual explanation for potential failures missed by standard metrics.

Significance and Claims
The paper claims that TRACE-Seg3D establishes counterfactual context auditing as a practical route toward transparent and reliable 3D medical image segmentation. By moving beyond simple overlap metrics, the framework allows clinicians to assess why a prediction might be unreliable (e.g., due to scanner variation) and whether the predicted tumor subregions are anatomically plausible. The authors emphasize that this approach improves robustness under institutional shifts without sacrificing in-distribution accuracy, offering a pathway to "trustworthy AI" where the model's dependence on nuisance cues is explicitly measured and mitigated. The work is modest in scope, focusing on MRI institutional shifts and using proxy-guided latents rather than claiming fully identifiable causal variables.