RADA: Region-Aware Dual-encoder Auxiliary learning for Barely-supervised Medical Image Segmentation

Shuang Zeng; Boxu Xie; Lei Zhu; Xinliang Zhang; Jiakui Hu; Zhengjian Yao; Yuanwei Li; Yuxing Lu

cs.CVApr 13, 2026

RADA: Region-Aware Dual-encoder Auxiliary learning for Barely-supervised Medical Image Segmentation

Authors: Shuang Zeng, Boxu Xie, Lei Zhu, Xinliang Zhang, Jiakui Hu, Zhengjian Yao, Yuanwei Li, Yuxing Lu,

Abstract

Deep learning has greatly advanced medical image segmentation, but its success relies heavily on fully supervised learning, which requires dense annotations that are costly and time-consuming for 3D volumetric scans. Barely-supervised learning reduces annotation burden by using only a few labeled slices per volume. Existing methods typically propagate sparse annotations to unlabeled slices through geometric continuity to generate pseudo-labels, but this strategy lacks semantic understanding, often resulting in low-quality pseudo-labels. Furthermore, medical image segmentation is inherently a pixel-level visual understanding task, where accuracy fundamentally depends on the quality of local, fine-grained visual features. Inspired by this, we propose RADA, a novel Region-Aware Dual-encoder Auxiliary learning pipeline which introduces a dual-encoder framework pre-trained on Alpha-CLIP to extract fine-grained, region-specific visual features from the original images and limited annotations. The framework combines image-level fine-grained visual features with text-level semantic guidance, providing region-aware semantic supervision that bridges image-level semantics and pixel-level segmentation. Integrated into a triple-view training framework, RADA achieves SOTA performance under extremely sparse annotation settings on LA2018, KiTS19 and LiTS, demonstrating robust generalization across diverse datasets.

Figures & tables

Fig. 1: Conceptual comparison of different prompt-based learning paradigms for barely-supervised medical image segmentation. The diagram illustrates three distinct approaches applied to sparse annotation datasets. From left to right: (1) DeSCO (No Prompt) relies solely on geometric propagation without semantic guidance; (2) SGTC (Text Prompt) incorporates textual semantics through class-specific prompts; (3) Our proposed RADA (Multimodal Prompt) integrates complementary modalities – combining text prompts with original image data and sparse annotation masks to achieve region-aware visual understanding.

Fig. 2: The pipeline of our proposed RADA framework. Specifically, a frozen vision encoder extracts region-aware visual features from the paired original images and their corresponding masks while simultaneously a frozen text encoder processes medical-domain prompts to generate semantic embeddings, with both modalities fused through an Adapter module for cross-modal alignment. The segmentation stage employs three parallel 3D sub-networks corresponding to orthogonal views, each incorporating the fused features via auxiliary learning to enhance decoder predictions through cross-modal guidance. The final output generates refined segmentation masks through coordinated multi-view predictions, utilizing registration-based pseudo-labels to expand limited supervision during training, thereby effectively leveraging sparse annotations through region-aware feature learning and multi-modal integration for precise segmentation.

Method	Labeled Slices	Scans Used		Metrics
Method	Labeled Slices	Labeled	Unlabeled	Dice $\uparrow$	Jaccard $\uparrow$	HD $\downarrow$	ASD $\downarrow$
MT [ 3 ]	SA	8	72	0.7365 $\pm$ 0.0801	0.5895 $\pm$ 0.1039	36.79 $\pm$ 10.09	9.42 $\pm$ 3.07
UA-MT [ 4 ]	SA	8	72	0.6615 $\pm$ 0.0912	0.5011 $\pm$ 0.1055	44.27 $\pm$ 7.11	17.66 $\pm$ 3.84
SASSNet [ 12 ]	SA	8	72	0.7767 $\pm$ 0.1582	0.6532 $\pm$ 0.1501	23.73 $\pm$ 13.22	5.70 $\pm$ 3.19
CPS [ 5 ]	SA	8	72	0.7859 $\pm$ 0.0976	0.6571 $\pm$ 0.1219	27.44 $\pm$ 9.48	6.39 $\pm$ 2.53
DTC [ 13 ]	SA	8	72	0.7975 $\pm$ 0.1452	0.6806 $\pm$ 0.1526	17.84 $\pm$ 8.89	2.69 $\pm$ 1.32

TABLE I: Quantitative comparisons with 8 SOTA methods on LA2018 dataset under 10% labeled cases. bold values denote the best performace. S, C, A indicate sagittal, coronal and axial annotated slices.

Method	Labeled Slices	Scans Used		Metrics
Method	Labeled Slices	Labeled	Unlabeled	Dice $\uparrow$	Jaccard $\uparrow$	HD $\downarrow$	ASD $\downarrow$
MT [ 3 ]	SA	19	171	0.8324 $\pm$ 0.1022	0.7255 $\pm$ 0.1438	31.86 $\pm$ 13.92	6.79 $\pm$ 4.48
UA-MT [ 4 ]	SA	19	171	0.8303 $\pm$ 0.0717	0.7160 $\pm$ 0.1052	30.58 $\pm$ 7.78	9.38 $\pm$ 3.41
SASSNet [ 12 ]	SA	19	171	0.8737 $\pm$ 0.0362	0.7775 $\pm$ 0.0574	23.70 $\pm$ 5.88	6.33 $\pm$ 1.79
CPS [ 5 ]	SA	19	171	0.8335 $\pm$ 0.0510	0.7177 $\pm$ 0.0722	47.24 $\pm$ 7.02	12.51 $\pm$ 2.51
DTC [ 13 ]	SA	19	171	0.6747 $\pm$ 0.0507	0.5114 $\pm$ 0.0615	63.33 $\pm$ 6.55	0.51 $\pm$ 0.19

TABLE II: Quantitative comparisons with 8 SOTA methods on KiTS19 dataset under 10% labeled cases.

Method	Labeled Slices	Scans Used		Metrics
Method	Labeled Slices	Labeled	Unlabeled	Dice $\uparrow$	Jaccard $\uparrow$	HD $\downarrow$	ASD $\downarrow$
MT [ 3 ]	SA	10	90	0.8990 $\pm$ 0.0341	0.8183 $\pm$ 0.0564	21.94 $\pm$ 12.64	3.73 $\pm$ 1.43
UA-MT [ 4 ]	SA	10	90	0.6336 $\pm$ 0.0677	0.4671 $\pm$ 0.0706	61.82 $\pm$ 11.60	25.52 $\pm$ 5.92
SASSNet [ 12 ]	SA	10	90	0.6408 $\pm$ 0.0701	0.4751 $\pm$ 0.0738	76.94 $\pm$ 13.55	31.70 $\pm$ 5.99
CPS [ 5 ]	SA	10	90	0.8935 $\pm$ 0.0494	0.8110 $\pm$ 0.0768	10.04 $\pm$ 8.37	2.18 $\pm$ 1.38
DTC [ 13 ]	SA	10	90	0.9206 $\pm$ 0.0408	0.8554 $\pm$ 0.0673	6.26 $\pm$ 5.75	1.65 $\pm$ 1.32

TABLE III: Quantitative comparisons with 8 SOTA methods on LiTS dataset under 10% labeled cases.

Method	Settings			Metrics
Method	Image Prompt	Mask Prompt	Registration	Dice $\uparrow$	Jaccard $\uparrow$	HD $\downarrow$	ASD $\downarrow$
SGTC (Baseline)	$\times$	$\times$	$\times$	0.8429 $\pm$ 0.0444	0.7309 $\pm$ 0.0647	19.91 $\pm$ 13.92	4.42 $\pm$ 3.66
SGTC + CLIP Vision Encoder	$\checkmark$	$\times$	$\times$	0.8504 $\pm$ 0.0382	0.7416 $\pm$ 0.0578	17.24 $\pm$ 12.44	4.43 $\pm$ 3.66
SGTC + Alpha-CLIP Vision Encoder	$\checkmark$	$\checkmark$	$\times$	0.8600 $\pm$ 0.0368	0.7561 $\pm$ 0.0561	18.23 $\pm$ 13.80	4.55 $\pm$ 3.92
RADA (Ours)	$\checkmark$	$\checkmark$	$\checkmark$	0.8652 $\pm$ 0.0282	0.7634 $\pm$ 0.0441	15.17 $\pm$ 10.90	3.24 $\pm$ 2.23

TABLE IV: Ablation Study of each component on LA2018 dataset.

Method	Settings			Metrics
Method	Image Prompt	Mask Prompt	Registration	Dice $\uparrow$	Jaccard $\uparrow$	HD $\downarrow$	ASD $\downarrow$
SGTC (Baseline)	$\times$	$\times$	$\times$	0.9254 $\pm$ 0.0349	0.8629 $\pm$ 0.0589	8.54 $\pm$ 11.33	2.56 $\pm$ 2.46
SGTC + CLIP Vision Encoder	$\checkmark$	$\times$	$\times$	0.9381 $\pm$ 0.0269	0.8846 $\pm$ 0.0464	10.08 $\pm$ 14.92	2.67 $\pm$ 2.48
SGTC + Alpha-CLIP Vision Encoder	$\checkmark$	$\checkmark$	$\times$	0.9392 $\pm$ 0.0241	0.8862 $\pm$ 0.0420	4.59 $\pm$ 8.04	1.58 $\pm$ 1.54
RADA (Ours)	$\checkmark$	$\checkmark$	$\checkmark$	0.9448 $\pm$ 0.0138	0.8957 $\pm$ 0.0245	3.09 $\pm$ 2.43	1.30 $\pm$ 0.98

TABLE V: Ablation Study of each component on KiTS19 dataset.

Method	Settings			Metrics
Method	Image Prompt	Mask Prompt	Registration	Dice $\uparrow$	Jaccard $\uparrow$	HD $\downarrow$	ASD $\downarrow$
SGTC (Baseline)	$\times$	$\times$	$\times$	0.9256 $\pm$ 0.0337	0.8631 $\pm$ 0.0548	10.08 $\pm$ 10.67	3.02 $\pm$ 2.19
SGTC + CLIP Vision Encoder	$\checkmark$	$\times$	$\times$	0.9309 $\pm$ 0.0337	0.8725 $\pm$ 0.0557	8.66 $\pm$ 11.29	2.61 $\pm$ 2.45
SGTC + Alpha-CLIP Vision Encoder	$\checkmark$	$\checkmark$	$\times$	0.9319 $\pm$ 0.0312	0.8739 $\pm$ 0.0526	7.74 $\pm$ 7.82	2.39 $\pm$ 1.88
RADA (Ours)	$\checkmark$	$\checkmark$	$\checkmark$	0.9363 $\pm$ 0.0203	0.8809 $\pm$ 0.0348	5.45 $\pm$ 6.53	1.78 $\pm$ 1.49

TABLE VI: Ablation Study of each component on LITS dataset.

Text Prompts	LA2018		KiTS19		LiTS
	Dice $\uparrow$	Jaccard $\uparrow$	Dice $\uparrow$	Jaccard $\uparrow$	Dice $\uparrow$	Jaccard $\uparrow$
None.	0.8573 $\pm$ 0.0313	0.7515 $\pm$ 0.0479	0.9401 $\pm$ 0.0138	0.8872 $\pm$ 0.0244	0.9302 $\pm$ 0.0246	0.8705 $\pm$ 0.0417
A photo of a [cls].	0.8518 $\pm$ 0.0376	0.7435 $\pm$ 0.0553	0.9408 $\pm$ 0.0159	0.8887 $\pm$ 0.0283	0.9336 $\pm$ 0.0217	0.8761 $\pm$ 0.0371
There is a [cls] in this computerized tomography/ magnetic resonance imaging.	0.8554 $\pm$ 0.0298	0.7485 $\pm$ 0.0462	0.9409 $\pm$ 0.0149	0.8890 $\pm$ 0.0264	0.9346 $\pm$ 0.0209	0.8780 $\pm$ 0.0358
An image containing the [cls], with the rest being background.	0.8652 $\pm$ 0.0282	0.7634 $\pm$ 0.0441	0.9448 $\pm$ 0.0138	0.8957 $\pm$ 0.0245	0.9363 $\pm$ 0.0203	0.8809 $\pm$ 0.0348

TABLE VII: Ablation Study of text prompts in Alpha-CLIP on LA2018, KiTS19 and LiTS dataset.

Fig. 3: Quantitative comparisons on KiTS19 dataset with 10% labeled case.

Fig. 4: Analysis of dynamic parameter $\beta$ .

Fig. 5: Ablation study of visual prompt configurations on LA2018, KiTS19, and LiTS datasets under 10% labeled cases. Red bars denote the optimal view strategy for each dataset.

Fig. 6: Visualization of Grad-CAM heatmaps on LA2018, KiTS19, LiTS dataset.

Labeled Slices	Image Encoder Slices	Metrics
Labeled Slices	Image Encoder Slices	Dice $\uparrow$	Jaccard $\uparrow$	HD $\downarrow$	ASD $\downarrow$
SSS	SS	0.7014 $\pm$ 0.0950	0.5479 $\pm$ 0.1134	38.35 $\pm$ 9.79	13.18 $\pm$ 4.70
CCC	CC	0.7387 $\pm$ 0.0692	0.5900 $\pm$ 0.0843	29.37 $\pm$ 11.67	9.76 $\pm$ 4.12
AAA	AA	0.7937 $\pm$ 0.0473	0.6604 $\pm$ 0.0644	27.63 $\pm$ 7.92	4.66 $\pm$ 4.27
ACA	AA	0.7859 $\pm$ 0.0492	0.6500 $\pm$ 0.0675	30.69 $\pm$ 10.43	5.91 $\pm$ 4.66
	CC	0.7909 $\pm$ 0.0448	0.6562 $\pm$ 0.0601	28.08 $\pm$ 9.33	4.09 $\pm$ 2.75

TABLE VIII: A comprehensive ablation study evaluating the impact of different slice selection strategies on segmentation performance of LA2018 dataset under 10% annotations.

Fig. 7: Quantitative comparisons on LA2018 dataset with 10% labeled case.

Fig. 8: Quantitative comparisons on LiTS dataset with 10% labeled case.

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 788 papers

SemiGDA: Generative Dual-distribution Alignment for Semi-Supervised Medical Image Segmentation
Apr 25, 2026Kaiwen Huang, Yi Zhou, Yizhe Zhang +2Distribution Matching Semi-Supervised Medical Image Segmentation
Nanjing University of Science and Technology, Nanjing, China · Southeast University, Nanjing, China
SSS: Semi-Supervised SAM-2 with Efficient Prompting for Medical Imaging Segmentation
Jun 10, 2025Hongjie Zhu, Xiwei Liu, Rundong Xue +5Promptable Image Segmentation Semi-Supervised Medical Image Segmentation
SWUN · MBZUAI · XJTU +2
Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations
Oct 10, 2026Jie Gui, Hang Tu, Yunfeng Dou +3Text-Guided Image Segmentation Promptable Image Segmentation
Key Laboratory of Intelligent Computing and Signal Processing, Ministry of Education, Anhui University, Hefei, China · School of Artificial Intelligence, Anhui University, Hefei, China · School of Computer Science and Technology, Anhui University, Hefei, China +1

Apr 25, 2026J/K move · Enter open · S save

SemiGDA: Generative Dual-distribution Alignment for Semi-Supervised Medical Image Segmentation

Kaiwen Huang, Yi Zhou, Yizhe Zhang, Jingxiong Li+1

Nanjing University of Science and Technology, Nanjing, China · Southeast University, Nanjing, China

Distribution Matching Semi-Supervised Medical Image Segmentation Medical Image Segmentation Semi-Supervised Learning Medical Image Representation Learning

Open paper AskPDF arXiv

Semi-supervised learning addresses label scarcity and high annotation costs in medical image segmentation by exploiting the latent information in unlabeled data to enhance model performance. Traditional discriminative segmentation relies on segmentation masks, neglecting feature-level distribution constraints. This limits robust semantic representation learning and adaptive modeling of unlabeled data in scenarios with few labels. To address these limitations, we propose SemiGDA, a novel Generative Dual-distribution Alignment framework for semi-supervised medical image segmentation. Our SemiGDA overcomes the reliance of discriminative methods on large labeled datasets by aligning feature and semantic distributions to boost semantic learning and scene adaptability. Specifically, we propose a Dual-distribution Alignment Module (DAM), which employs two structurally distinct encoders to model image and mask feature distributions. It enforces their alignment in the latent space via distributional constraints, establishing structured feature consistency. Moreover, we design a Consistency-Driven Skip Adapter (CDSA) strategy, which introduces dual skip adapters (Image and Mask) to fuse multi-scale features via skip connections. Using a consistency loss, CDSA enhances cross-branch semantic alignment and reinforces fine-grained semantic consistency. Experimental results on diverse medical datasets show that our method outperforms other state-of-the-art semi-supervised segmentation methods. Code is released at: https://github.com/taozh2017/SemiGDA.

Abstract

Medical Image Segmentation

Explore similar work

SemiGDA: Generative Dual-distribution Alignment for Semi-Supervised Medical Image Segmentation

SSS: Semi-Supervised SAM-2 with Efficient Prompting for Medical Imaging Segmentation

Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations