Unsupervised Domain Adaptation with Target-Only Margin Disparity Discrepancy

TL;DR

A novel unsupervised domain adaptation (UDA) framework based on the formalism of Margin Disparity Discrepancy (MDD), which improves target domain performance through a reformulation of the original MDD optimization framework is proposed, which achieves state-of-the-art performance in UDA, as well as in the few-shot setting.

Abstract

In interventional radiology, Cone-Beam Computed Tomography (CBCT) is a helpful imaging modality that provides guidance to practicians during minimally invasive procedures. CBCT differs from traditional Computed Tomography (CT) due to its limited reconstructed field of view, specific artefacts, and the intra-arterial administration of contrast medium. While CT benefits from abundant publicly available annotated datasets, interventional CBCT data remain scarce and largely unannotated, with existing datasets focused primarily on radiotherapy applications. To address this limitation, we leverage a proprietary collection of unannotated interventional CBCT scans in conjunction with annotated CT data, employing domain adaptation techniques to bridge the modality gap and enhance liver segmentation performance on CBCT. We propose a novel unsupervised domain adaptation (UDA) framework based on the formalism of Margin Disparity Discrepancy (MDD), which improves target domain performance through a reformulation of the original MDD optimization framework. Experimental results on CT and CBCT datasets for liver segmentation demonstrate that our method achieves state-of-the-art performance in UDA, as well as in the few-shot setting.

Figures (4)

• Figure 1: Axial slices from abdominal CT (left, source) and CBCT (right, target) of the same patient using identical visualization windows. The liver is shown as a blue overlay. High-intensity regions within the liver caused by intra-arterial contrast enhancement in CBCT are marked with yellow arrows.
• Figure 2: Our feature-alignment UDA for segmentation. A U-Net is decomposed into a feature extractor $\psi$, and a segmentation head $f$. An adversary $f'$ is built as a duplication of $f$ for adversarial training during UDA. After UDA, the features $z^S = \psi(x^S)$ and $z^T = \psi(x^T)$ are supposed to be domain-invariant, and $f'$ is removed at inference. The loss $\mathcal{L}_{task}$ is the supervision segmentation loss, and $\mathcal{L}_{CE}$ is used for UDA.
• Figure 3: Segmentation results from various 3D methods. Ground truth mask is shown in red, and predicted segmentations in blue. High-intensity regions within the liver that cause the networks to fail are marked with yellow arrows. The same axial slice is displayed for each method with the corresponding prediction overlaid. The SAM-MED 3D prediction was generated using 5 randomly sampled points. The F1 score above each method, reported as a percentage, represents the 3D F1 score computed across the entire volume.
• Figure 4: Stability Analysis of Hyperparameters in 2D CT to CBCT Unsupervised Domain Adaptation (UDA). This figure reports the F1 scores obtained on the target domain test set for different combinations of the hyperparameters $\alpha$ and $\gamma$, with F1 values presented as percentages. The consistent performance across varying values of $\alpha$ and $\gamma$ confirms the robustness and stability of our method with respect to these hyperparameters.