BARL: Bilateral Alignment in Representation and Label Spaces for Semi-Supervised Volumetric Medical Image Segmentation

TL;DR

This work tackles the challenge of semi-supervised volumetric medical image segmentation under limited annotations by proposing BARL, a dual-space framework that enforces alignment in both representation and label spaces. It introduces region-level and lesion-instance representation alignment alongside label-space modules DPR and PCBC within a co-training setup, unifying constraints across multiple decoder scales. Extensive experiments across BraTS 2020/2021/2023 MEN, CBCT Tooth, and IXI demonstrate that BARL achieves state-of-the-art performance under 5–20% labeling, with strong Dice and boundary metrics and robust cross-dataset generalization. Ablation studies validate the contribution of each component, including the importance of instance-level representation alignment, multi-level DPR, and uncertainty-guided PCBC, establishing BARL as a powerful baseline for data-efficient 3D medical image segmentation.

Abstract

Semi-supervised medical image segmentation (SSMIS) seeks to match fully supervised performance while sharply reducing annotation cost. Mainstream SSMIS methods rely on \emph{label-space consistency}, yet they overlook the equally critical \emph{representation-space alignment}. Without harmonizing latent features, models struggle to learn representations that are both discriminative and spatially coherent. To this end, we introduce \textbf{Bilateral Alignment in Representation and Label spaces (BARL)}, a unified framework that couples two collaborative branches and enforces alignment in both spaces. For label-space alignment, inspired by co-training and multi-scale decoding, we devise \textbf{Dual-Path Regularization (DPR)} and \textbf{Progressively Cognitive Bias Correction (PCBC)} to impose fine-grained cross-branch consistency while mitigating error accumulation from coarse to fine scales. For representation-space alignment, we conduct region-level and lesion-instance matching between branches, explicitly capturing the fragmented, complex pathological patterns common in medical imagery. Extensive experiments on four public benchmarks and a proprietary CBCT dataset demonstrate that BARL consistently surpasses state-of-the-art SSMIS methods. Ablative studies further validate the contribution of each component. Code will be released soon.

Figures (5)

• Figure 1: Typical alignment protocols and lesion fragmentation statistics.(a) Label space. Soft logits are aligned by Kullback--Leibler or mean-squared error, whereas hard pseudo-labels (via arg max) are aligned by Dice or cross-entropy. (b) Representation space. We locate class-specific feature vectors, then enforce alignment to widen inter-class margins while tightening intra-class clusters. Fragmentation in Medical Images.Top: A natural scene with its category-level mask; pixels of each class coalesce into a single, compact connected component. Bottom: A medical mask. Left—foreground (white) vs. background (black); right—lesions encoded by colour, where voxels of the same class split into multiple disconnected fragments.
• Figure 2: Overview of our proposed BARL strategy, built upon a co-training architecture. The strategy enforces constraints in two dimensions: the representation space, where region-level and instance-level feature vectors are obtained and consistency operations are applied; and the label space, where the proposed DPR module constrains outputs among multi-layer decoders and the PCBC module corrects inconsistent results under the co-training framework.
• Figure 3: A summary of mainstream semi-supervised learning architectures, including Mean Teacher style, FixMatch-style consistency regularization, Co-Training with mutual pseudo labeling, and Singleton-style frameworks.
• Figure 4: Performance under different combinations of representation spaces and dimensions on BraTs2020 in 20% label ratio.
• Figure 5: External data validation on BraTs2021 dataset using well-trained model on BraTs2020 dataset within 10% ratio.