Mutual Learning-Driven Multi-Scale Feature Fusion Network for Medical Image Segmentation

Transformer-based encoder networks (TransUNet) have emerged as a prominent research focus in medical image segmentation. To address the limitations of existing approaches that insufficiently capture local contextual relationships between adjacent patches during Transformer encoder integration and lack inter-channel information interaction in decoder operations, this study proposes a Mutual Learning Multi-Scale Feature Fusion model (MLMSFF). The framework employs parallel network training with three principal innovations: Firstly, a multi-scale fusion module enhances global contextual awareness through hierarchical feature aggregation. Secondly, an Importance Mapping Distillation (IMD) module facilitates cross-network knowledge transfer of critical anatomical details, complemented by Kullback-Leibler divergence (KLD) regularization to align feature distributions across networks. Thirdly, a hybrid loss function combines symmetrical cross-entropy for class imbalance mitigation with Dice loss for boundary constraint. Comprehensive evaluations demonstrate MLMSFF’s superior performance: On the MM-WHS dataset, it achieves mean Dice scores of $\mathbf{8 9. 7 6 \%}$ (CT) and $84.65 \%$ (MRI) with reduced ASSD of 3.07 mm and 2.46 mm respectively. For MS-CMR MRI data, the model attains $86.76 \%$ Dice score and 0.82 mm ASSD. Comparative analysis with TransUNet, Att-UNet, and other UNet variants reveals that MLMSFF achieves enhanced segmentation accuracy for both internal structures and boundary regions while maintaining comparable parameter budgets.