Deep Learning Models for Automatic Diagnosis and Efficacy Assessment of Scars: A Case Study

Scars affect millions of patients each year. High recurrence makes it difficult to treat. There are a variety of treatments for different subtypes of scars; however, there are significant interindividual differences in the efficacy of these treatments. Thus, an accurate and objective tool to assist physicians in evaluating scars is necessary. This study aimed to demonstrate the effectiveness and application value of intelligent systems in scar clinical diagnosis and efficacy assessment. An intelligent system can be designed to assist physicians in providing personalized treatment plans for scarring patients and monitoring the effectiveness of treatment. In this study, deep learning models were trained using standardized imaging of 3000 normal skin images, 2142 scar images and 1431 other skin diseases images. All scar images were matched to Vancouver Scar Scale (VSS) scores, as determined by three experienced dermatologists, to train efficacy assessment models. We evaluated the performance of various models and compared the best-performing model with five dermatologists. The results showed that Swin Transformer achieved the highest Area Under the Curve (AUC) of 0.9696 in scar subtype classification, the precision and recall were 0.9027 and 0.8876, respectively. For scar efficacy assessment, Swin Transformer also showed the best consistency with the lowest average Mean Absolute Error (MAE, 0.5137) for VSS item scores. Compared with human performance, Swin Transformer showed better objectivity and coherence (average MAE of 0.4369) but the evaluation was based on a very small sample. The study demonstrated that the deep learning system enabled more accurate scar subtype diagnosis and more objective and consistent efficacy assessment, demonstrating the potential to assist physicians in personalizing scar management.

The online version contains supplementary material available at https://doi.org/10.1007/s43657-025-00251-8.