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

Jul 16, 2026

Upload by Eden

Medical image segmentation models often score highly on benchmarks yet remain vulnerable to scanner, protocol, and institutional appearance differences. These context shifts change MRI appearance without altering the underlying lesion, enabling models to rely on nuisance cues that overlap metrics like Dice and HD95 do not reveal. TRACE-Seg3D introduces counterfactual context auditing for 3D glioma segmentation, preserving lesion-relevant evidence while systematically varying imaging context to measure stability and anatomical plausibility for case-level reliability assessment beyond overlap-based evaluation.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages16

Reading Time40 min

Introduction

Reliability limits of overlap metrics

Causal view and counterfactual reliability

TRACE-Seg3D framework

Audit evidence: lesion support, context sensitivity, plausibility

Experiments and contributions

arXiv :2607 .07038v 1 [ cs .CV] 8 Jul 2026 TRACE-Seg3D: Counterfactual Context Auditing For Robust 3D Glioma Segmentation Under Institutional Shift Nguyen Linh Dan Le 1, Nguyen Pham Hoang Le2, and Tran Dang Khoi3 1 The University of Melbourne, Melbourne, Australia 2 University of Information Technology, Vietnam National University, Vietnam 3 Industrial University of Ho Chi Minh City, Vietnam [dan.le@ieee.org](dan.le@ieee.org) , [22520982@gm.uit.edu.vn](22520982@gm.uit.edu.vn) , [khoitran08102006@gmail.com](khoitran08102006@gmail.com) Abstract. Medical image segmentation models can achieve strong benchmark performance while remaining sensitive to scanner, protocol, and institutional variation. These context shifts alter image appearance without changing the underlying lesion, allowing models to exploit nuisance cues that Dice and HD95 fail to expose. We present TRACE-Seg3D, acounterfactual context auditing framework for robust 3D medical image segmentation. TRACE-Seg3D preserves lesion-relevant evidence and systematically varies imaging context to quantify prediction stability under controlled context shifts. The framework pairs each segmentation with audit evidence for context sensitivity and anatomical plausibility, enabling case-level reliability assessment beyond overlap-based evaluation. Experiments on BraTS and UTSW glioma segmentation benchmarks demonstrate competitive in-distribution and cross-domain performance. TRACE-Seg3D also exposes context-sensitive failure modes missed by conventional metrics. These results establish counterfactual context auditing as a practical route toward transparent and reliable 3D medical image segmentation under distribution shift. Our code is available at TRACE-Seg3D repository. Keywords: 3D Medical Image Segmentation · Counterfactual Context Auditing · Trustworthy AI · Distribution Shift · Robust Segmentation 1 Introduction Segmentation models may achieve strong Dice scores on benchmark datasets yet fail silently when deployed on images acquired from a different scanner or institution. Although the predicted mask changes substantially, conventional overlap metrics provide little insight into why the prediction fails or whether the new prediction can be trusted. Medical image segmentation aims to delineate clinically meaningful structures from medical images and has become a core component of quantitative diagnosis, treatment planning, and longitudinal disease monitoring. In glioma MRI, the segmentation target is structured rather than a single foreground region. Whole tumor (WT), tumor core (TC), and enhancing tumor 2 N.L.D. Le et al. (ET) describe related but clinically distinct subregions, with ET being particularly sensitive because it is often small, visually ambiguous, and important for assessing active disease. Modern 3D segmentation models, including U-Net variants, self-configuring pipelines, transformer-based encoders, and ConvNeXtstyle volumetric networks, have advanced this task by optimizing overlap and boundary metrics [7–10, 17 , 19 , 21 , 22] . Although these models provide accurate masks on standard validation splits, overlap metrics offer limited insight into whether a prediction remains reliable when imaging conditions change. Scanner hardware, acquisition parameters, reconstruction pipelines, and institutional protocols can alter MRI appearance without changing the underlying tumor. As a result, a model may learn a hidden dependence between site-specific appearance and the predicted mask. This dependence is difficult to diagnose from Dice or HD95 alone: the same score cannot distinguish lesion-grounded prediction from context-driven prediction, nor can it reveal whether the predicted ET, TC, and WT regions obey their anatomical hierarchy. This observation motivates a causal view of trustworthy 3D tumor segmentation. We treat disease evidence as the mechanism that should determine the anatomical mask, and imaging context as a mechanism th

FAQs

Why can a 3D glioma segmentation model fail under institutional shift even with strong Dice scores?

Scanner hardware, acquisition parameters, and reconstruction pipelines can alter MRI appearance while the underlying tumor evidence stays the same. Overlap metrics such as Dice and HD95 cannot distinguish lesion-grounded predictions from context-driven ones, so failures may go unnoticed.

What is the key idea behind TRACE-Seg3D’s counterfactual context auditing?

TRACE-Seg3D treats disease evidence as the mechanism that should determine the anatomical mask, while imaging context changes the observed image without directly forcing the segmentation once disease evidence is fixed. By varying context under a counterfactual setup, the method assesses stability and robustness.

How does TRACE-Seg3D improve auditability beyond overlap-based evaluation?

TRACE-Seg3D pairs each segmentation with audit evidence covering lesion support, context sensitivity, and anatomical plausibility. This enables case-level reliability assessment and helps reveal context-sensitive failure modes that conventional metrics miss.

Jul 16, 2026

Upload by Eden