DCSCSA U-Net: A Deeper and Compact Spatial-Channel Synergistic Attention U-Net for Scar Segmentation and Detection

Scars are common outcomes of the skin wound healing process and have significant impacts on patients’ appearance and physiological functions; thus, accurate segmentation and detection are of great significance. Currently, scar assessment is mainly performed by visual inspection, which lacks precision. In this study, a segmentation network based on an encoder-decoder architecture, namely the Deeper and Compact Spatial-Channel Synergistic Attention U-Net (DCSCSA U-Net), is proposed for automatic segmentation and detection of scars. The model is built upon DCSAU-Net and incorporates a multi-dimensional SCSA mechanism to achieve multi-scale feature fusion and enhancement of key regions. Five-fold cross-validation experiments were conducted on both the ISIC 2018 and locally collected HPSH datasets, and superior IoU and DSC scores were achieved compared to existing benchmark models, reaching 0.832±0.29 and 0.879 ± 0.28 (ISIC 2018), and 0.541 ± 0.23 and 0.662 ± 0.25 (HPSH), respectively. Furthermore, due to the limited sample size of the HPSH dataset, a transfer learning strategy was adopted. After pre-training on ISIC 2018 and transferring to HPSH, IoU and DSC were further improved by 9.4% and 6.6%, reaching 0.635±0.18 and 0.728±0.24, respectively. In summary, DCSCSA U-Net demonstrates excellent generalization and segmentation performance, providing an effective technical solution for automatic scar and medical image segmentation and detection.