Exploring the influence of transformer-based multimodal modeling on clinicians’ diagnosis of skin diseases: A quantitative analysis

Abstract

Objectives The study aimed to propose a multimodal model that incorporates both macroscopic and microscopic images and analyze its influence on clinicians’ decision-making with different levels of experience. Methods First, we constructed a multimodal dataset for five skin disorders. Next, we trained unimodal models on three different types of images and selected the best-performing models as the base learners. Then, we used a soft voting strategy to create the multimodal model. Finally, 12 clinicians were divided into three groups, with each group including one director dermatologist, one dermatologist-in-charge, one resident dermatologist, and one general practitioner. They were asked to diagnose the skin disorders in four unaided situations (macroscopic images only, dermatopathological images only, macroscopic and dermatopathological images, all images and metadata), and three aided situations (macroscopic images with model 1 aid, dermatopathological images with model 2&3 aid, all images with multimodal model 4 aid). The clinicians’ diagnosis accuracy and time for each diagnosis were recorded. Results Among the trained models, the vision transformer (ViT) achieved the best performance, with accuracies of 0.8636, 0.9545, 0.9673, and AUCs of 0.9823, 0.9952, 0.9989 on the training set, respectively. However, on the external validation set, they only achieved accuracies of 0.70, 0.90, and 0.94, respectively. The multimodal model performed well compared to the unimodal models, achieving an accuracy of 0.98 on the external validation set. The results of logit regression analysis indicate that all models are helpful to clinicians in making diagnostic decisions [Odds Ratios (OR) > 1], while metadata does not provide assistance to clinicians (OR < 1). Linear analysis results indicate that metadata significantly increases clinicians’ diagnosis time ( P < 0.05), while model assistance does not ( P > 0.05). Conclusions The results of this study suggest that the multimodal model effectively improves clinicians’ diagnostic performance without significantly increasing the diagnostic time. However, further large-scale prospective studies are necessary.