BCMDA: Bidirectional Correlation Maps Domain Adaptation for Mixed Domain Semi-Supervised Medical Image Segmentation

Bentao Song; Jun Huang; Qingfeng Wang

cs.CVMar 25, 2026

BCMDA: Bidirectional Correlation Maps Domain Adaptation for Mixed Domain Semi-Supervised Medical Image Segmentation

Authors: Bentao Song, Jun Huang, Qingfeng Wang

Organizations: School of Computer Science and Technology, Southwest University of Science and Technology, Mianyang, Sichuan, 621010, China

Abstract

In mixed domain semi-supervised medical image segmentation (MiDSS), achieving superior performance under domain shift and limited annotations is challenging. This scenario presents two primary issues: (1) distributional differences between labeled and unlabeled data hinder effective knowledge transfer, and (2) inefficient learning from unlabeled data causes severe confirmation bias. In this paper, we propose the bidirectional correlation maps domain adaptation (BCMDA) framework to overcome these issues. On the one hand, we employ knowledge transfer via virtual domain bridging (KTVDB) to facilitate cross-domain learning. First, to construct a distribution-aligned virtual domain, we leverage bidirectional correlation maps between labeled and unlabeled data to synthesize both labeled and unlabeled images, which are then mixed with the original images to generate virtual images using two strategies, a fixed ratio and a progressive dynamic MixUp. Next, dual bidirectional CutMix is used to enable initial knowledge transfer within the fixed virtual domain and gradual knowledge transfer from the dynamically transitioning labeled domain to the real unlabeled domains. On the other hand, to alleviate confirmation bias, we adopt prototypical alignment and pseudo label correction (PAPLC), which utilizes learnable prototype cosine similarity classifiers for bidirectional prototype alignment between the virtual and real domains, yielding smoother and more compact feature representations. Finally, we use prototypical pseudo label correction to generate more reliable pseudo labels. Empirical evaluations on three public multi-domain datasets demonstrate the superiority of our method, particularly showing excellent performance even with very limited labeled samples. Code available at https://github.com/pascalcpp/BCMDA.

Figures & tables

Figure 1: From (a) to (d), the divisions of the training and testing sets represent different scenario settings, with the difficulty increasing sequentially.

Figure 2: Illustration of distribution differences under MiDSS: (a) the original distribution of labeled and unlabeled data, (b) labeled data distribution after a unidirectional transformation used in SymGD, and (c) the well-aligned virtual domain distribution formed by both labeled and unlabeled data after the bidirectional transformation with our method.

Figure 3: This figure illustrates the overall architecture of bidirectional correlation maps domain adaptation (BCMDA). (a) The specific process of virtual data generation, where CIS stands for the correlation map-informed image synthesis operation. (b) Mixed training data is generated using dual bidirectional CutMix, enabling efficient knowledge transfer by leveraging the virtual domain as a bridge, where AVG refers to the operation of averaging two probability maps. (c) The prototypical alignment and pseudo label correction (PAPLC) component, which employs bidirectional prototype alignment for smooth feature aggregation. Additionally, for classification results erroneously assigned to the class on the right due to a linear decision boundary, it is observed that their features are closer to the prototype on the left. In such cases, the predictions are corrected to the left-side class.

Figure 4: Illustration of the bidirectional correlation maps, where the green border and blue border denote the labeled and unlabeled images, respectively, and the correlation map in the middle represents $\mathcal{C}^{x^{w}\_u^{w}}$ . Here, we ignore the softmax operation, so $\mathcal{C}^{u^{w}\_x^{w}}$ is simply the transpose of $\mathcal{C}^{x^{w}\_u^{w}}$ . From the figure, we can see that each column shows the correlations between all labeled pixels and a single unlabeled pixel, while each row shows the correlations between all unlabeled pixels and a single labeled pixel.

Figure 5: This figure compares BCMix applied to raw images with the domain gap and to the distribution-aligned virtual images. On the left, BCMix is applied to real domain images, where the pasted patches exhibit stylistic differences from the backgrounds, hindering knowledge transfer. On the right, BCMix operates on virtual domain images, enabling smooth transitions between the patches and backgrounds for near seamless integration and enhanced knowledge transfer.

Datasets	Configure	Domain1	Domain2	Domain3	Domain4	Domain5	Domain6
Fundus	Name	Domain1	Domain2	Domain3	Domain4	/	/
Fundus	Train/Test Count	50/51	99/60	320/80	320/80	/	/
Prostate	Name	RUNMC	BMC	HCRUDB	UCL	BIDMC	HK
Prostate	Train/Test Count	338/83	305/79	373/95	133/42	225/36	136/22
M&Ms	Name	Vendor A	Vendor B	Vendor C	Vendor D	/	/
M&Ms	Train/Test Count	1030/274	1342/325	525/129	550/135	/	/

Table 1: Detailed information of all datasets.

Methods	#L	(Optic Cup / Optic Disc Segmentation) Dice $\uparrow$				Dice $\uparrow$	Jaccard $\uparrow$	95HD $\downarrow$	ASD $\downarrow$	$p$ value (Dice)
Methods	#L	Domain 1	Domain 2	Domain 3	Domain 4	Avg.				$p$ value (Dice)
U-Net	20	57.92 / 71.80	64.98 / 70.86	48.88 / 65.03	41.52 / 65.41	$60.80_{\pm 1.17}$	$51.99_{\pm 0.93}$	$47.62_{\pm 0.93}$	$29.28_{\pm 0.60}$	$<0.01$
Upper bound		85.47 / 93.38	80.60 / 90.94	85.28 / 93.02	85.65 / 93.24	$88.45_{\pm 0.02}$	$80.32_{\pm 0.04}$	$7.42_{\pm 0.01}$	$3.73_{\pm 0.04}$	$<0.01$
UA-MT ( Yu et al., 2019 )	20	63.53 / 79.63	67.60 / 79.69	45.32 / 56.69	41.63 / 60.45	$61.82_{\pm 7.89}$	$52.31_{\pm 7.50}$	$42.47_{\pm 8.73}$	$25.79_{\pm 7.86}$	$<0.01$
SIFA ( Chen et al., 2020 )	20	59.89 / 75.98	66.76 / 84.47	64.47 / 85.31	57.24 / 73.84	$70.99_{\pm 4.54}$	$58.48_{\pm 5.25}$	$18.59_{\pm 2.22}$	$11.76_{\pm 1.17}$	$<0.01$
FixMatch ( Sohn et al., 2020 )	20	81.32 / 91.49	72.64 / 87.17	78.72 / 92.78	74.66 / 88.70	$83.44_{\pm 0.06}$	$73.76_{\pm 0.39}$	$11.35_{\pm 0.59}$	$5.61_{\pm 0.01}$	$<0.01$

Table 2: Comparison with other methods on the Fundus dataset, where #L denotes the number of labeled samples. The best results are highlighted in bold , while the second-best results are underlined .

Figure 6: Visualization results on the Fundus dataset.

Methods	#L	(Prostate Segmentation) Dice $\uparrow$						Dice $\uparrow$	Jaccard $\uparrow$	95HD $\downarrow$	ASD $\downarrow$	$p$ value (Dice)
Methods	#L	RUNMC	BMC	HCRUDB	UCL	BIDMC	HK	Avg.				$p$ value (Dice)
U-Net	20	20.79	19.02	17.74	15.25	18.91	29.12	$20.14_{\pm 0.45}$	$14.82_{\pm 1.30}$	$121.66_{\pm 6.20}$	$82.41_{\pm 3.10}$	$<0.01$
U-Net	40	28.71	35.61	21.23	43.56	20.60	27.62	$29.56_{\pm 1.62}$	$24.03_{\pm 1.12}$	$97.97_{\pm 4.04}$	$65.18_{\pm 0.94}$	$<0.01$
Upper bound		88.56	88.56	85.60	88.63	88.90	89.44	$88.28_{\pm 0.05}$	$80.67_{\pm 0.05}$	$10.07_{\pm 0.04}$	$4.13_{\pm 0.02}$	$<0.01$
UA-MT ( Yu et al., 2019 )	20	18.44	15.17	16.15	33.54	14.32	13.50	$18.52_{\pm 0.55}$	$13.21_{\pm 0.34}$	$129.30_{\pm 2.26}$	$87.65_{\pm 2.80}$	$<0.01$
FixMatch ( Sohn et al., 2020 )	20	82.93	64.24	54.11	70.23	11.15	82.82	$60.91_{\pm 0.89}$	$53.22_{\pm 0.73}$	$46.61_{\pm 4.02}$	$26.85_{\pm 2.13}$	$<0.01$

Table 3: Comparison with other methods on the Prostate dataset.

Figure 7: Visualization results on the Prostate dataset using 20 labeled data.

Method	#L	(LV / MYO / RV Segmentation) Dice $\uparrow$				Dice $\uparrow$	Jaccard $\uparrow$	95HD $\downarrow$	ASD $\downarrow$	$p$ value (Dice)
Method	#L	Vendor A	Vendor B	Vendor C	Vendor D	Avg.				$p$ value (Dice)
U-Net	5	33.87 / 17.97 / 27.00	48.14 / 49.23 / 29.02	16.22 / 13.37 / 19.18	43.20 / 34.06 / 35.37	$30.56_{\pm 1.28}$	$22.19_{\pm 3.51}$	$68.44_{\pm 3.64}$	$44.32_{\pm 6.02}$	$<0.01$
U-Net	20	41.67 / 27.89 / 28.39	56.56 / 49.59 / 47.19	55.89 / 45.64 / 42.60	69.31 / 58.30 / 57.52	$48.38_{\pm 6.55}$	$40.61_{\pm 5.22}$	$43.02_{\pm 7.00}$	$28.68_{\pm 8.20}$	$<0.01$
Upper bound		91.19 / 83.65 / 84.00	92.48 / 85.56 / 84.54	92.26 / 84.35 / 84.27	91.82 / 82.90 / 83.50	$86.71_{\pm 0.15}$	$78.52_{\pm 0.17}$	$4.41_{\pm 0.13}$	$2.24_{\pm 0.13}$	$<0.01$
UA-MT ( Yu et al., 2019 )	5	16.32 / 8.85 / 10.54	38.10 / 36.26 / 16.47	19.29 / 13.89 / 12.13	33.26 / 21.99 / 16.00	$20.26_{\pm 0.51}$	$14.73_{\pm 0.59}$	$77.47_{\pm 1.17}$	$54.19_{\pm 0.82}$	$<0.01$
FixMatch ( Sohn et al., 2020 )	5	80.67 / 66.81 / 57.78	85.09 / 80.19 / 75.35	83.18 / 75.71 / 67.40	89.60 / 78.44 / 79.04	$76.61_{\pm 1.69}$	$66.85_{\pm 0.81}$	$\underline{9.65}_{\pm 0.80}$	$5.12_{\pm 0.36}$	$<0.01$

Table 4: Comparison with other methods on the M&Ms dataset.

Figure 8: Visualization results on the M&Ms dataset using 20 labeled data.

Methods	Avg. Dice $\uparrow$	Avg. ASD $\downarrow$	$p$ value (Dice)	Dice of Each Class
Methods	Avg. Dice $\uparrow$	Avg. ASD $\downarrow$	$p$ value (Dice)	Sp	RK	LK	Ga	Es	Li	St	Ao	IVC	PSV	PA	RAG	LAG
V-Net (fully)	$62.09_{\pm 1.2}$	$10.28_{\pm 3.9}$	$<0.01$	84.6	77.2	73.8	73.3	38.2	94.6	68.4	72.1	71.2	58.2	48.5	17.9	29.0
Sup Only	$23.99_{\pm 7.3}$	$74.26_{\pm 5.9}$	$<0.01$	25.0	27.1	9.3	4.4	43.4	85.2	54.0	48.8	14.7	0.1	0.0	0.0	0.0
UA-MT ( Yu et al., 2019 )	$20.26_{\pm 2.2}$	$71.67_{\pm 7.4}$	$<0.01$	48.2	31.7	22.2	0.0	0.0	81.2	29.1	23.3	27.5	0.0	0.0	0.0	0.0
URPC ( Luo et al., 2022b )	$25.68_{\pm 5.1}$	$72.74_{\pm 15.5}$	$<0.01$	66.7	38.2	56.8	0.0	0.0	85.3	33.9	33.1	14.8	0.0	5.1	0.0	0.0
CPS ( Chen et al., 2021 )	$33.55_{\pm 3.7}$	$41.21_{\pm 9.1}$	$<0.01$	62.8	55.2	45.4	35.9	0.0	91.1	31.3	41.9	49.2	8.8	14.5	0.0	0.0

Table 5: Comparison with other methods on the Synapse dataset.

Figure 9: Visualization results on the Synapse dataset.

Figure 10: Performance visualization comparison of BCMDA with other methods on the Fundus and Prostate datasets using 20 labeled data.

Method	BCMix	FixMix	PDMix	FDA	AVG	PA	BPA	PPLC	(Optic Cup / Optic Disc Segmentation) Dice $\uparrow$				Dice $\uparrow$	Jaccard $\uparrow$	95HD $\downarrow$	ASD $\downarrow$
Method	BCMix	FixMix	PDMix	FDA	AVG	PA	BPA	PPLC	Domain 1	Domain 2	Domain 3	Domain 4	Avg.
#1	✓								82.76 / 92.51	73.75 / 89.50	81.91 / 92.62	81.73 / 93.09	85.98	76.85	9.41	4.44
#2	✓	✓							82.91 / 93.48	79.66 / 91.66	83.49 / 93.19	84.38 / 93.54	87.79	79.33	7.76	3.70
#3	✓		✓						82.68 / 93.35	79.41 / 90.92	83.30 / 93.18	83.46 / 93.83	87.51	78.95	7.86	3.78
#4	✓	✓	✓						83.76 / 93.52	81.52 / 91.83	83.96 / 93.60	84.60 / 94.14	88.37	80.12	7.56	3.59
#5	✓	✓	✓	✓					81.82 / 93.39	80.74 / 90.22	83.54 / 93.29	83.32 / 93.52	87.48	78.77	7.83	3.81

Table 6: Ablation study results on Fundus dataset.

$W^{\prime}$	Dice $\uparrow$	Jaccard $\uparrow$	95HD $\downarrow$	ASD $\downarrow$
$W^{\prime}$	Avg.
$\frac{1}{2}W$	88.84	80.86	7.27	3.47
$\frac{1}{4}W$	88.95	81.03	7.21	3.45
$\frac{1}{8}W$	88.90	80.93	7.38	3.45

Table 7: Comparison with different

W^{\prime}

Figure 11: Hyper-parameter analysis of $\lambda_{\text{fix}}$ , $\alpha$ , and $\tau_{\text{temp}}$ .

Method	Performance		Complexity
Method	Dice $\uparrow$	ASD $\downarrow$	Epoch(s)	Iteration(s)	Memory(GB)
SymGD	87.20	3.92	101.31	0.20	4.07
BCMDA	88.95	3.45	70.88	0.14	4.79

Table 8: Comparison of performance and complexity for different methods

Figure 14: This figure illustrates generated virtual images for challenging samples where misinterpretation by the model occurs.

Explore similar work

Jul 5, 2026

Beyond Random Sampling: Distribution-Aware Alignment for Semi-Supervised Medical Image Segmentation

Precise medical image segmentation is crucial for clinical diagnosis and treatment planning, yet relies heavily on expensive expert annotations. Semi-supervised medical image segmentation (SSMIS) offers a cost-effective solution but typically operates under the assumption of independent and identically distributed (i.i.d.) data, defaulting to random sampling. While statistically valid at scale, this strategy suffers from severe representation bias in low-data regimes, failing to capture the heterogeneous medical data manifold. To address this, we propose a highly data-efficient framework driven by distribution alignment. First, we introduce an offline Distribution-Aware Sample Selection strategy. By leveraging Vision Foundation Models (VFMs) and our designed Density-K-Center algorithm, we explicitly identify representative structural anchors, establishing a more representative labeled domain. Second, to bridge the remaining distribution gap, we propose the Memory-guided Copy-Paste (MCP) module. Tailored for the inherent class imbalance in medical scans, MCP leverages a semantic memory mechanism to retrieve historically consistent priors for cross-domain alignment, encouraging semantic consistency. Coupled with an easy-to-hard progressive schedule, this framework effectively mitigates early-stage pseudo-label noise. Extensive experiments on six diverse 2D and 3D datasets demonstrate strong segmentation performance, particularly in extremely low-labeled scenarios (\eg, 1/16 ratio).

Weihao Yan, Yeqiang Qian, Yi Dong +1

School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, Shanghai, China · Department of Ultrasound, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China

Data Augmentation Data Selection Distribution Matching Semi-Supervised Medical Image Segmentation

Ask PDF

Apr 25, 2026

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

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.

Kaiwen Huang, Yi Zhou, Yizhe Zhang +2

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

Ask PDF

Jun 19, 2026

One Size does not Fit All: Heterogeneous Latent Space Alignment for Unsupervised Domain Adaptation

Domain shift remains a major obstacle to the reliable deployment of machine learning models in high-stakes environments such as healthcare. While Domain adaptation aims to mitigate these effects, existing approaches suffer from limited expressiveness of latent representations and a reliance on handcrafted, static augmentations. In this work, we address these limitations by proposing a novel deep learning architecture for Unsupervised Domain Adaptation (UDA), specifically optimized for medical image segmentation. Our framework, ADualVUOT, integrates a dual-encoder Variational Autoencoder (VAE) with Continuous Normalizing Flows (CNFs) to increase modeling flexibility and posterior expressiveness. To achieve domain alignment, we leverage Unbalanced Optimal Transport (UOT) through the Gaussian-Gromov-Wasserstein (GGW) distance, which handles structural and topological discrepancies between domains. Furthermore, we incorporate an adversarial augmentation scheme to synthesize worst-case compositions, thus enhancing model robustness. Extensive experiments on medical imaging benchmarks show significant gains over prior OT-based approaches.

Evangelia Koskinioti, Yi Shen, Georgios Stamou +1

Department of Electrical and Computer Engineering, Duke University, NC, USA · Department of Mechanical Engineering, Duke University, NC, USA · Department of Electrical and Computer Engineering, National Technical University of Athens, Athens, Greece.

Unsupervised Domain Adaptation Domain Adaptation Medical Image Segmentation

Ask PDF