Enhanced Multi-Modal Heart Segmentation Approach based on Swin-Transformer for Small Datasets

Multi-modal heart segmentation is a crucial and challenging task in medical image analysis, playing a pivotal role in disease diagnosis and treatment. While most current research on ventricular segmentation relies on ample data under ideal conditions, the risk of overfitting arises when applied to small-scale datasets. This paper proposes an enhanced multi-modal ventricular segmentation method based on the U-Net network and vision Transformer model to address limitations of both the full convolutional network, which struggles with learning long-distance feature relationships, and the traditional vision Transformer model with its high computational load. Additionally, a boundary loss function is introduced. Through testing and validation on the MM-WHS dataset and comparison with various methods, significant improvements have been achieved in MRI and CT image segmentation. This not only aids in understanding the strengths and weaknesses of different models for medical image segmentation tasks but also offers insights for optimizing models and selecting suitable ones for small-scale datasets.