TwinTrack: Post-hoc Multi-Rater Calibration for Medical Image Segmentation

2026-04-17

AI Summary

This work addresses the challenge of inconsistent expert annotations in segmenting pancreatic ductal adenocarcinoma (PDAC) on contrast-enhanced CT scans, a problem exacerbated by conventional deep learning approaches that assume a single ground truth, leading to poorly calibrated probabilistic outputs and limited interpretability. To overcome this, the authors propose TwinTrack, a novel framework that introduces posterior calibration into medical image segmentation for the first time. TwinTrack maps ensemble model outputs to the Mean Human Response (MHR)—the average of multiple expert annotations—enabling predicted probabilities to directly reflect inter-expert uncertainty. Requiring only a small number of multiply-annotated samples for calibration, the method achieves state-of-the-art performance on the MICCAI 2025 CURVAS-PDACVI benchmark, significantly enhancing both the reliability and clinical interpretability of segmentation outputs.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Segmentation
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Security and Privacy: Tracking, profiling, and countermeasures against them
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Pancreatic ductal adenocarcinoma (PDAC) segmentation on contrast-enhanced CT is inherently ambiguous: inter-rater disagreement among experts reflects genuine uncertainty rather than annotation noise. Standard deep learning approaches assume a single ground truth, producing probabilistic outputs that can be poorly calibrated and difficult to interpret under such ambiguity. We present TwinTrack, a framework that addresses this gap through post-hoc calibration of ensemble segmentation probabilities to the empirical mean human response (MHR) -the fraction of expert annotators labeling a voxel as tumor. Calibrated probabilities are thus directly interpretable as the expected proportion of annotators assigning the tumor label, explicitly modeling inter-rater disagreement. The proposed post-hoc calibration procedure is simple and requires only a small multi-rater calibration set. It consistently improves calibration metrics over standard approaches when evaluated on the MICCAI 2025 CURVAS-PDACVI multi-rater benchmark.

Problem

• medical image segmentation
• inter-rater disagreement
• uncertainty
• calibration
• pancreatic ductal adenocarcinoma

Innovation

• post-hoc calibration
• multi-rater disagreement
• medical image segmentation
• probabilistic interpretation
• ensemble segmentation