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

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 (14)

• 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.