Research on cross-modal generalization of medical image segmentation based on ResUNet optimization

Abstract

Due to fundamental differences in imaging physics, signal reconstruction, and noise characteristics, medical images from different modalities exhibit severe cross-modal heterogeneity, manifesting as inconsistent intensity distributions, varying lesion morphology, and modality-specific noise. These differences lead to poor generalization of traditional segmentation models and inaccurate delineation of blurred lesion boundaries. To address these issues, this paper proposes an improved ResUNet for cross-modal medical image segmentation. ResNet50 is adopted as the encoder backbone, combined with a Feature Pyramid Network (FPN) to construct multi-scale feature hierarchies. A Squeeze-and- Excitation (SE) module is integrated to suppress background noise and enhance lesion-related features. A dynamic feature recalibration module is designed to adaptively adjust multi-modal feature fusion weights, and cross-modal contrastive learning is introduced to reduce inter-modal feature distribution discrepancies. Extensive experiments on three mainstream datasets, including ISIC2018 (dermoscopy), BUSI (ultrasound), and LiTS2017 (CT), demonstrate that the proposed model outperforms state-of-the-art models such as TransUNet and Swin-Unet. The method achieves a favorable balance between segmentation accuracy and computational efficiency, providing reliable technical support for multi-modal clinical diagnosis.