Differentiating scalp psoriasis from seborrheic dermatitis (SD) by visual inspection is challenging, often requiring dermoscopy. To enhance the diagnostic precision of dermoscopic results, we aimed to explore the accuracy (ACC) and clinical feasibility of using deep learning algorithms to differentiate dermoscopic images of scalp psoriasis from SD. We collected 1955 dermoscopic images with a clinical diagnosis of scalp psoriasis and SD. We trained and validated convolutional neural network (CNN)–based network (HIENet) and three transformer‐based networks: vision transformer (ViT), data‐efficient image transformer (DeiT) and cross‐attention multiscale visual transformer (CrossViT) using fivefold cross‐validation. In the fivefold cross‐validation, ViT achieved a diagnostic ACC of 97.09% with an area‐under‐the‐curve value of 0.9960, outperforming HIENet, DeiT, and slightly surpassing CrossViT. In the external validation dataset, ViT achieved a 94% ACC, which was higher than that of Doctor 1 and Doctor 2, whereas Doctor 3, with a 98% ACC, performed better. The Transformer‐based ViT model demonstrated good performance and potential in differentiating dermoscopic images of scalp psoriasis from SD. It was able to improve the ACC and efficiency of dermoscopic diagnosis.
Lai et al. (Thu,) studied this question.