Authors - Sumet Jirattisak, Tanatorn Tanantong, Nittaya Chemkomnerd Abstract - Psoriasis is a chronic autoimmune skin disease, and accurate diagnosis remains challenging due to the shortage of dermatologists and the subjective na ture of visual assessment. To address this challenge, this study developed an au tomated classification system using three deep learning architectures, Efficient Net-B4, MobileNetV3, and Vision Transformer, within a transfer learning frame work to classify Psoriasis, Healthy Skin, and Psoriasis-like Disorder images. The models were fine-tuned and evaluated using 5-fold cross-validation on three da tasets: the Thammasat University Hospital dataset, the Kaggle dataset, and a combined dataset derived from DermNet and a previously published study in volving Indian patients. EfficientNet-B4 achieved the highest accuracy on the TUH dataset (99.68%) and the Dermnet-India dataset (94.40%), while Mo bileNetV3 performed best on the Kaggle dataset (96.88%) and required the short est training time. Overall, the results show that EfficientNet-B4 offers superior predictive performance, whereas MobileNetV3 provides a better balance be tween accuracy and computational efficiency. The findings confirm that transfer learning is a time-efficient approach for psoriasis classification, reducing training time and computational cost while maintaining acceptable performance, particu larly under limited clinical data conditions.
