A Multilayer Skin Disease Classification on SkinBench Dataset Scans

Abstract

A significant issue in the clinical dermatology is still the diagnosis of skin diseases in regions where the dermatologists and diagnostic devices cannot be easily accessed. Many skin diseases are similar in appearance, and therefore, they are not properly diagnosed even among the long-term experienced clinicians. The existing deep learning diagnostic solutions only prove a small number of diseases and represent the disease classification problem as a one-step issue, which might reduce their usefulness in practice regarding a wide range of different types of diseases. To address these issues, the current paper introduces the SkinBench dataset and the elaborated multi-layered skin diseases classification system that leads to hierarchical recognition of three levels (L1: Normal vs. Abnormal), (L2)- seven-class (L3) -subclass classification of Eczema, Fungal and Pox disease families. A collection of 9 major skin disease types were generated that were optimized and further subclassified into groups to enable hierarchical reasoning.ResNet50, DenseNet121, MobileNetV3, EfficientNet-B0, VGG16, VGG19, a custom CNN and hybrid SwinDenseNet were subsequently trained and tested on the same experimental pipeline with hyperparameters and class-balancing schemes being optimized to be robust to variations in the characteristics of skin color, lighting, and lesions. The quantitative results demonstrate that the suggested multilayer design is highly superior to conventional single-stage classification schemes, L1 module is highly accurate in the process of normal and abnormal image classification, strong routing to L2-L3, disease classification accuracy at L2, and specific subclass prediction at L3, which is paramount in the process of clinical results interpretation. SwinDenseNet and DenseNet121 had higher scores of macro-F1 and hierarchical architecture reduced the confusion of classes because of the form of decision boundaries in models as defined by confusion matrices, ROC curves, precision-recall curves, and modelbased comparison tables. Another application, a Streamlit-based deployment dashboard was developed to enable real-time inference with automated L1USD -USD L2USD L3 routing and comprehensible subclass explanations, which will have a tremendous future as a clinical pre-screening tool and a telehealth dermatology one. Altogether, SkinBench dataset and the proposed multilayer framework can be deemed as a scalable, user-friendly, and high-performance solution to the multilayer skin disease diagnosis that is an invaluable addition to the credible AI-assisted dermatological care.