Abstract
Semi-supervised learning offers a compelling path to accurate pancreas segmentation with limited annotations, but its success hinges on the quality of pseudo-labels, a double-edged sword where reliable predictions propagate useful supervision while erroneous ones misguide the network and compound over training. The most problematic region is the pancreas boundary, where low CT contrast and anatomical variability make predictions inherently uncertain and errors systematically patterned rather than incidental. The Translation Consistent Co-training (TraCoCo) framework achieves strong performance by reducing spatial-context memorisation through co-training and translation consistency, but it does not specifically address boundary regions. On labeled data, its standard Dice and CE losses treat all voxels equally, ignoring systematic boundary errors. On unlabeled data, its confidence filtering is applied uniformly without boundary-aware consideration, discarding uncertain predictions entirely. We introduce three complementary extensions to TraCoCo: Optimal Transport-based Correction in Disagreement Region (OT-CDR) augments per-voxel error with a Wasserstein distance to correct holistic boundary patterns; Confidence-Adaptive Pseudo-Label Gating (CA-PLG) replaces hard confidence thresholds with continuous entropy-based weighting; and Texture-Guided Alignment (TGA) grounds both corrections in CT gradient evidence. On the NIH Pancreas benchmark with only 20% labelled data, our method achieves 82.50% Dice, outperforming the state-of-the-art methods. Consistent improvements on the MSD Pancreas dataset further demonstrate that explicitly addressing boundary regions during both supervised and semi-supervised training enhances the effectiveness of semi-supervised pancreas segmentation. Our code is available at: https://github.com/abdora-ai/OT-CDR-Pancreas-SSL
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_022.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=nupK2LlIOf
BibTex
@InProceedings{SalAmi_Distributional_MICCAISAT2026,
author = { Saleknia, Amir Hossein AND Kheyrkhah, Alireza AND Azad, Reza AND Houshmand, Sina AND Bagci, Ulas AND Sulaiman, Alaa},
title = { { Distributional Mutual Correction for Semi-Supervised Pancreas Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}