TwinTrack: Calibrated PDAC CT Segmentation

About

Editor:

 Tryzis 

Summary

TwinTrack is a pancreatic ductal adenocarcinoma (PDAC) segmentation algorithm for contrast-enhanced CT that was designed for ambiguous multi-rater settings, where disagreement between experts reflects genuine uncertainty rather than simple annotation noise. The method combines a coarse-to-fine segmentation pipeline with a post-hoc calibration step that maps predicted tumor probabilities to the mean human response (MHR), i.e. the fraction of expert annotators labeling a voxel as tumor. This makes the output easier to interpret in practice: calibrated probabilities can be read as the expected proportion of experts who would assign the tumor label at a given voxel.

If you use TwinTrack in your research, please cite:

@misc{kirscher2026twintrack,
      title={TwinTrack: Post-hoc Multi-Rater Calibration for Medical Image Segmentation}, 
      author={Tristan Kirscher and Alexandra Ertl and Klaus Maier-Hein and Xavier Coubez and Philippe Meyer and Sylvain Faisan},
      year={2026},
      eprint={2604.15950},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2604.15950}, 
}

Mechanism

Target population. TwinTrack targets patients with pancreatic ductal adenocarcinoma (PDAC) on contrast-enhanced CT.

Algorithm description. The method uses a two-stage segmentation pipeline followed by multi-rater-aware post-hoc calibration:

1. A low-resolution nnU-Net first localizes the pancreas on the full CT and defines a high-recall region of interest (ROI) by dilation.
2. Inside this ROI, an ensemble of 3 independently trained high-resolution nnU-Nets refines the prediction.
3. The ensemble outputs are averaged and merged into a binary tumor vs. non-tumor prediction.
4. A post-hoc isotonic regression calibration model is then applied to the tumor probabilities so that they align with the voxel-wise mean human response (MHR) observed in a small multi-rater calibration set.

Inputs and outputs.

• Input: contrast-enhanced CT volume.
• Output: calibrated voxel-wise PDAC probability map and corresponding tumor segmentation.
• Interpretation of the probability map: values close to 1 indicate strong expected agreement among expert annotators that the voxel belongs to tumor; intermediate values reflect regions of expert disagreement or boundary ambiguity.

Interfaces

This algorithm implements all of the following input-output combinations:

Validation and Performance

The segmentation models were trained on PANORAMA batch 4. Multi-rater annotations were used only for post-hoc calibration: in the paper, the calibration mapping was fit on the CURVAS-PDACVI training split (40 CT scans with 5 expert annotations) without retraining the base segmenter. Reported results below are bootstrap means on the CURVAS-PDACVI test set (n = 64).

Ambiguity-aware segmentation and calibration metrics¶

Challenge Performance

Uses and Directions

TwinTrack was developed for research use in uncertainty-aware PDAC segmentation on contrast-enhanced CT, especially in settings where multiple expert annotations are available or clinically relevant disagreement is expected.

Appropriate uses include:

• generating calibrated PDAC probability maps that explicitly reflect inter-rater ambiguity,
• supporting ambiguity-aware segmentation studies,
• supporting downstream analyses in which uncertainty-aware tumor extent is useful, such as vascular invasion assessment derived from the predicted tumor extent.

TwinTrack should not be used as a standalone diagnostic system, should not replace expert radiological review or manual delineation, and has not been validated in this work for other organs, tumor types, imaging modalities, or non-contrast CT acquisitions.

Warnings

• The calibrated output represents an estimate of expert agreement, not a direct probability of malignancy, pathology confirmation, or clinical outcome.
• Calibration improves the interpretability of the score distribution, but it does not guarantee correct localization if the underlying segmenter fails.
• Performance may degrade under dataset shift, including different scanners, contrast phases, reconstruction settings, institutions, annotation protocols, or patient populations.
• Boundary regions with intermediate probabilities should be interpreted as ambiguous areas, not as labeling errors.
• This algorithm was evaluated in a challenge/research setting and should be used accordingly.

Common Error Messages