Leveraging Labelled Data Knowledge: A Cooperative Rectification Learning Network for Semi-supervised 3D Medical Image Segmentation

TL;DR

This work tackles semi-supervised 3D medical image segmentation with limited labelled data by introducing Cooperative Rectification Learning Network (CRLN), Dynamic Interaction Module (DIM), and Collaborative Positive Supervision (CPS). CRLN learns multiple class prototypes from labelled data and uses DIM to compute a holistic relationship map $\mathscr{M}(x)$ that adaptively rectifies pseudo-labels, enabling more unlabelled data to contribute during training. CPS provides a moderated contrastive learning signal for uncertain regions by forming unassertive positives that blend prototype-driven and mean representations, improving discrimination at boundaries and low-contrast areas. The approach yields state-of-the-art performance across LA, Pancreas-CT, and BraTS19 datasets, reduces reliance on annotated data, and demonstrates robust improvements in pseudo-label quality and segmentation of challenging regions, with broader implications for reducing annotation burdens in clinical pipelines.

Abstract

Semi-supervised 3D medical image segmentation aims to achieve accurate segmentation using few labelled data and numerous unlabelled data. The main challenge in the design of semi-supervised learning methods consists in the effective use of the unlabelled data for training. A promising solution consists of ensuring consistent predictions across different views of the data, where the efficacy of this strategy depends on the accuracy of the pseudo-labels generated by the model for this consistency learning strategy. In this paper, we introduce a new methodology to produce high-quality pseudo-labels for a consistency learning strategy to address semi-supervised 3D medical image segmentation. The methodology has three important contributions. The first contribution is the Cooperative Rectification Learning Network (CRLN) that learns multiple prototypes per class to be used as external knowledge priors to adaptively rectify pseudo-labels at the voxel level. The second contribution consists of the Dynamic Interaction Module (DIM) to facilitate pairwise and cross-class interactions between prototypes and multi-resolution image features, enabling the production of accurate voxel-level clues for pseudo-label rectification. The third contribution is the Cooperative Positive Supervision (CPS), which optimises uncertain representations to align with unassertive representations of their class distributions, improving the model's accuracy in classifying uncertain regions. Extensive experiments on three public 3D medical segmentation datasets demonstrate the effectiveness and superiority of our semi-supervised learning method.

Figures (15)

• Figure 1: Quality of pseudo-labels as a function of training iteration under the 10% partition semi-supervised learning protocol on the LA dataset. (a) Proportion of reliable predictions within all pseudo-labels before (green curve) and after (red curve) our proposed rectification. (b) Segmentation accuracy (Dice) results of all pseudo-labels (orange), reliable pseudo labels (green), all rectified pseudo labels (blue), and reliable rectified pseudo labels (red). During training, especially in the early stages, only a small percentage of the predictions of the pseudo-labels are reliable, which produces relatively inaccurate segmentation. After applying our proposed rectification, the percentage of reliable pseudo-labels and their respective segmentation accuracy are significantly improved (Sec. \ref{['sec:Ablation_Study']} has more details).
• Figure 2: The main contributions of this paper are the Cooperative Rectification Learning Network (CRLN, yellow box), Dynamic Interaction Module (DIM, pink box), and the Collaborative Positive Supervision (CPS, blue box). CRLN learns multiple class-wise prototypes that work as knowledge priors for the rectification of pseudo-labels. DIM aims to acquire holistic relationships across multiple class prototypes and unlabelled data to rectify pseudo labels. CPS trains uncertain representations to get closer to their unassertive positive learning key, enabling the model to better discriminate such uncertain regions.
• Figure 3: The architecture of the proposed model. Based on the teacher-student structure, the Cooperative Rectification Learning Network (CRLN) consists of two stages. In the learning stage, multiple category prototypes are built and initialised. Subsequently, the Dynamic Interaction Module (DIM) implements pairwise interactions, as well as spatial-aware and cross-class aggregation between prototypes and the semantics of the labelled data to obtain the holistic relationship map $\mathscr{M}(x)$ which adaptively improves the segmentation quality of $\hat{y}$ with \ref{['eq:rectification_student']}. By minimising the deviation between predictions and labels, the proposed CRLN effectively learns valuable category prototypes and understands how to use them for voxel-level correction. In the rectification stage, the learned category prototypes serve as prior knowledge to rectify the pseudo-labels $\bar{y}$. After rectification, the higher-quality pseudo-labels $\bar{y}_{r}$ are used as supervision signals. Moreover, the Collaborative Positive Supervision (CPS) mechanism constructs unassertive centres by integrating learned category prototypes and category mean representations, allowing for better contrastive learning ($\ell_{cp}(\cdot)$ in \ref{['eq:cp']} of representations with lower predictive confidence in the student network. This empowers the model to distinguish uncertain regions.
• Figure 4: The architecture of the Dynamic Interaction Module (DIM).
• Figure 5: The architecture of the Collaborative Positive Supervision (CPS) mechanism.