A co-training approach integrating CNN and Mamba for semi-supervised 3D medical image segmentation

Abstract

Accurate 3D medical image segmentation is vital for clinical applications but hindered by the scarcity of expert-annotated data, posing challenges for fully supervised models that require extensive labels. To address this issue, semi-supervised learning (SSL) has emerged as an effective solution by enabling models to learn from limited labeled data and achieve performance comparable to fully supervised learning, thereby significantly reducing the burden of manual annotation. Among various SSL strategies, deep co-training has demonstrated its effectiveness, yet current methods suffer from excessive information sharing between subnetworks or convergence during the training process. In this paper, we propose a semi-supervised 3D medical image segmentation method based on heterogeneous co-training, which combines the local feature perception capability of Convolutional Neural Networks (CNNs) with the efficient sequence modeling ability of the Mamba architecture. This integration enables more robust and complementary feature learning. To further enhance the utilization of unlabeled data, we introduce a Confidence-Aware Consistency(CAC) loss, which enforces consistency between model predictions while maintaining the learning capacity of individual networks. In addition, we propose a TriMix data augmentation strategy and incorporate a feature perturbation mechanism to increase the diversity of training samples and improve generalization. Extensive experiments were performed on three publicly available datasets: BraTS2019, left atrium (LA) and pancreas. The proposed method achieved superior performance compared to existing semi-supervised segmentation approaches, as measured by Dice score and HD95, demonstrating its effectiveness in enhancing the accuracy and robustness of 3D medical image segmentation with limited labeled data.