Accurate medical image segmentation plays a critical role in clinical diagnosis, treatment planning, and a wide range of healthcare applications. Although U-shaped CNNs and Transformer-based architectures have shown promise, CNNs struggle to capture long-range dependencies, whereas Transformers suffer from quadratic growth in computational cost as image resolution increases. To address these issues, we propose HCMUNet, a novel medical image segmentation model that innovatively combines the local feature extraction capabilities of CNNs with the efficient long-range dependency modeling of Mamba, enhancing feature representation while reducing computational cost. In addition, HCMUNet features a redesigned skip connection and a novel attention module that integrates multi-scale features to recover spatial details lost during down-sampling and to promote richer cross-dimensional interactions. HCMUNet achieves Dice Similarity Coefficients (DSC) of 90.32%, 81.52%, and 92.11% on the ISIC 2018, Synapse multi-organ, and ACDC datasets, respectively, outperforming baseline methods by 0.65%, 1.05%, and 1.39%. Furthermore, HCMUNet consistently outperforms U-Net and Swin-UNet, achieving average Dice score improvements of approximately 5% and 2% across the evaluated datasets. These results collectively affirm the effectiveness and reliability of the proposed model across different segmentation tasks.

PurposeAccurate pixel-level segmentation is essential for medical image analysis, particularly in assisting diagnosis and treatment planning. However, fully supervised learning methods rely heavily on high-quality annotated data, which are often scarce due to the high cost of manual labeling, privacy concerns, and limited availability. We aim to reduce reliance on precise annotations and improve segmentation performance under weak supervision.ApproachWe propose scribble position and temporal contrast learning (SPTCL), an innovative segmentation method that combines contrastive learning with weak supervision. Our method leverages the spatial continuity in 3D medical image volumes and the anatomical similarities across different cardiac phases to construct a contrastive learning task for robust feature representation from unlabeled data. To enhance the feature extraction capabilities, we employ a pre-trained encoder, which is initially trained on the ACDC dataset using contrastive learning to capture robust feature representations. This pre-trained encoder is then transferred to a weakly supervised segmentation network with a dual-branch decoder for further fine-tuning on the task. The predictions from both branches are fused to generate refined pseudo-labels, which are iteratively used to guide network training with only coarse scribble annotations.ResultsExperiments on the ACDC dataset show that SPTCL outperforms existing models, achieving a Dice coefficient of 90.5%, with a 2.5% improvement over the baseline and a 1.7% improvement over the latest model. Furthermore, SPTCL reduces training time by ∼33%.ConclusionsSPTCL effectively addresses the challenges of limited annotation in medical image segmentation by uniting contrastive learning with weak supervision. It demonstrates strong potential for practical deployment in clinical settings where high-quality labels are difficult to obtain.