A Novel Cross-Scale Feature Enhancement and Fusion Method for Medical Image Segmentation

Transformer-based neural network techniques have proven highly effective in medical image segmentation tasks. However, their inherent limitations in handling scale-related information may result in an excessive focus on deep semantic details at the expense of essential contextual cues. This can lead to segmentation outputs with gaps and indistinct edges. To address this issue, we have introduced specific strategies to bolster multiscale information integration. Notably, we have devised the Channel-Spatial Interactive Module (CSIM) to promote the harmonious interplay and enrichment of spatial and channel details across different scales. This is followed by the fusion of these features with adjacent-scale counterparts, facilitating mutual guidance for multiscale feature acquisition. Ultimately, we employ a stepwise process of upsampling and information fusion at diverse scales. Our approach has undergone rigorous evaluation across multiple datasets, consistently outperforming alternative methods across a range of metrics.