Publication

Abstract : Erythrodermic psoriasis is an uncommon but severe type that requires prompt and precise evaluation to direct clinical treatment. It is characterized by widespread skin inflammation and a peeling rash across the entire body. This work presents a sophisticated erythema severity score model that uses transfer learning with EfficientNet-B0 and is tuned for high precision in clinical conditions. By utilizing the EfficientNet-B0 architecture that was previously trained on ImageNet, we refined the model to identify the degree of erythema in psoriasis lesions, classifying it into five categories: Mild, Absent, Moderate, Severe, and Very Severe. Deeper layers of the model were specifically designed for erythema qualities, while the early layers were frozen to maintain wide feature representations. A two-stage training method was employed to increase detection accuracy, first stabilizing the model with frozen layers and then fine-tuning selectively at a slower learning rate. The final model outperformed baseline models and demonstrated outstanding reliability with an accuracy of 0.918. These outcomes demonstrate the model’s potential for significant clinical application in assessing the severity of erythema, offering a practical and efficient approach for dermatology’s real-time severity grading.