Semi-supervised cardiac MRI image segmentation via learning consistency under differential perturbations

Problem: Fully-supervised learning-based methods have achieved remarkable success in cardiac MRI image segmentation due to the availability of large-scale labeled datasets. However, obtaining sufficient cardiac annotations not only requires expert participation, but is also costly and time-consuming, which limits the application of fully supervised learning methods. Aim: To reduce the dependence of the training of segmentation models on a large amount of labeled data, this study proposes an improved semi-supervised method based on UniMatch, which aims to achieve accurate segmentation of cardiac MRI images with limited labeled data. Method: Two modules named Differential Perturbation Pool and Consistency Regularization Module (DPPCRM) and Dual-dimensional Feature Perturbation Module (DFPM) are proposed and integrated into UniMatch to enhance its performance in cardiac MRI image segmentation tasks. In the DPPCRM, In DPPCRM, two differentiated perturbation pools are first designed to increase the relevance of the strongly perturbed images for the cardiac segmentation task and to explore a wider perturbation space. Then, a consistency regularization method is constructed in DPPCRM to mine and exploit the feature hidden in the strongly perturbed images, which forces the model to better extract features from samples generated by the newly designed perturbation pools. While in the DFPM, perturbations are performed on the features extracted from weakly perturbed images in both channel and spatial dimensions, which enriches the types of feature perturbations and allows the model to learn more robust feature representations. Results: Extensive experiments on the Automated Cardiac Diagnosis Challenge (ACDC) 2017 dataset and the Multi-Sequence Cardiac MR Segmentation Challenge (MSCMR) 2019 challenge dataset show that the proposed improved-UniMatch outperforms the state-of-the-art semi-supervised methods, which demonstrates its effectiveness for cardiac MRI image segmentation. On the ACDC dataset, the DSC and HD₉₅ obtained by proposed method with 5% and 10% labeled data are 0.880 and 4.025 mm, 0.898 and 2.527 mm, respectively. On the MSCMR dataset, the DSC and HD₉₅ obtained by proposed method with 5% and 10% labeled data are 0.726 and 12.688 mm, 0.748 and 8.706 mm, respectively. Furthermore, ablation studies prove that the above two modules are all effective and essential for proposed method to achieve this very superior performance. Conclusion: The proposed improved-UniMatch framework effectively addresses the limitations of the original UniMatch by introducing enhanced perturbation strategies and leveraging feature-level consistency. The results confirm that the proposed method can achieve excellent cardiac segmentation accuracy with limited labeled data.