MSF-TransUNet: A Multi-Scale Fusion Approach for Precise Cardiac Image Segmentation
Nuorui Zhang, Jingtao Li · 2024
Cardiac image segmentation is crucial in the detection and treatment of cardiovascular disorders. The complexity of cardiac anatomy, morphological variations under different pathological conditions, and inconsistencies in imaging quality pose significant challenges to existing medical image segmentation methods. In order to tackle these difficulties, this work presents the MSF-TransUNet, an innovative structure specifically developed for the purpose of segmenting cardiac images. This model harnesses the synergistic capabilities of Convolutional Neural Networks (CNNs) and Transformer architectures to tackle the intricacies of cardiac anatomy and variations in imaging conditions. The MSF-TransUNet incorporates an Efficient Pyramid Squeeze Attention Network (EPSANet) as its encoder, enhancing multi-scale feature representation and long-range channel dependencies. Additionally, a Multi-Scale Fusion Module (MSFM) is developed to integrate multi-scale features from various encoder layers within the skip connections, utilizing convolution operations and the SCSE attention mechanism for efficient feature fusion. To overcome the challenges of cardiac structural complexity and data imbalance, the model introduces an Adaptive t-vMF Dice Loss, dynamically adjusting similarity metrics to enhance segmentation performance across different categories and regions. Extensive evaluations on two widely used cardiac image segmentation datasets, CAMUS and ACDC, demonstrate that the MSF-TransUNet outperforms existing state-of-the-art models on several metrics. Notably, our model achieves average Dice coefficients of 0.9243 and 0.9206, showcasing its flexibility and accuracy in handling scenarios with poor image quality or complex cardiac anatomy.