[Paper] Unsupervised Domain Adaptation with Target-Only Margin Disparity Discrepancy

Source: arXiv - 2603.09932v1

Overview

This paper tackles a very practical problem in medical imaging: how to get accurate liver segmentations on interventional Cone‑Beam CT (CBCT) scans when only a handful of annotated CT images are available. By treating CBCT as an unlabeled target domain and leveraging a large, fully‑annotated CT dataset as the source, the authors introduce a new unsupervised domain adaptation (UDA) technique that pushes the performance of deep segmentation models far beyond what was previously possible.

Key Contributions

• Target‑only Margin Disparity Discrepancy (MDD‑T): a reformulated version of the classic Margin Disparity Discrepancy loss that relies solely on target‑domain predictions, simplifying optimization and improving stability.
• End‑to‑end UDA pipeline for liver segmentation that couples a standard encoder‑decoder network with the MDD‑T regularizer, requiring no target‑domain labels.
• Extensive empirical validation on a proprietary interventional CBCT dataset and a public CT liver segmentation benchmark, showing state‑of‑the‑art results in both pure UDA and few‑shot (1–5 labeled CBCT scans) scenarios.
• Demonstration of few‑shot synergy: the same framework can seamlessly incorporate a few annotated CBCT slices, yielding further gains without redesign.

Methodology

1. Base segmentation model – a conventional 3‑D U‑Net (or similar encoder‑decoder) trained on the source CT data with a standard Dice loss.
2. Domain discrepancy term – instead of the original MDD that needs both source and target predictions, the authors propose MDD‑T, which measures the margin between the most confident class and the second‑most confident class only on target samples. Intuitively, it encourages the model to make confident, well‑separated predictions on CBCT images, even though no ground‑truth masks are provided.
3. Adversarial‑style optimization – the training alternates between:
   • Feature extractor update: minimize the segmentation loss on CT and the MDD‑T loss on CBCT, pulling target features toward a decision boundary that yields high confidence.
   • Margin maximizer update: a small auxiliary network tries to increase the margin on CBCT, acting like a “discriminator” that pushes the feature extractor to produce clearer class separations.
4. Few‑shot extension – when a few labeled CBCT slices are available, a tiny Dice term is added for those samples, letting the model fine‑tune directly on the target domain.

The whole pipeline is trained end‑to‑end with standard stochastic gradient descent, requiring only the raw CBCT volumes (no preprocessing beyond typical intensity clipping).

Results & Findings

• The MDD‑T loss converges faster and is less sensitive to hyper‑parameter tuning than the original MDD.
• Visual inspection shows markedly reduced CBCT‑specific artefacts (e.g., streaks, limited field‑of‑view truncation) in the predicted liver masks.
• Ablation studies confirm that removing the margin term drops Dice back to ~0.78, highlighting its central role.

Practical Implications

• Rapid deployment in interventional suites – hospitals can now train a liver segmentation model on publicly available CT datasets and adapt it to their own CBCT scanners without costly manual annotation campaigns.
• Improved navigation and dose planning – accurate real‑time liver masks enable better guidance for catheter placement, lesion targeting, and radiation dose calculations during minimally invasive procedures.
• Generalizable framework – the target‑only MDD formulation can be swapped into any segmentation backbone (e.g., nnU‑Net, Transformers) and applied to other modality gaps (MRI ↔ CT, ultrasound ↔ CT).
• Cost‑effective few‑shot scaling – acquiring just a handful of annotated CBCT volumes yields near‑optimal performance, dramatically lowering the barrier for small clinics or research labs.

Limitations & Future Work

• Dataset specificity – the experiments use a proprietary CBCT collection from a single vendor; cross‑vendor robustness remains to be proven.
• 3‑D memory footprint – training the full 3‑D U‑Net with MDD‑T can be GPU‑intensive; future work could explore patch‑based or mixed‑precision strategies.
• Extension beyond liver – while liver segmentation is a solid proof‑of‑concept, adapting the method to more complex, multi‑organ or tumor segmentation tasks may require additional regularization.
• Theoretical analysis – the paper provides empirical justification for MDD‑T but a deeper theoretical understanding of why target‑only margins work so well would strengthen the contribution.

Overall, the proposed Target‑Only Margin Disparity Discrepancy offers a pragmatic, high‑impact route to bring state‑of‑the‑art deep segmentation into the interventional radiology workflow with minimal annotation overhead.

Authors

• Gauthier Miralles
• Loïc Le Folgoc
• Vincent Jugnon
• Pietro Gori

Paper Information

• arXiv ID: 2603.09932v1
• Categories: cs.CV
• Published: March 10, 2026
• PDF: Download PDF