Multiclass Dermatological Disease Classification Using Advanced Convolution Deep Learning Technique

Abstract

Skin diseases affect people of every age and region, and many of them can be treated well when they are found early. Diagnosis, however, depends on trained dermatologists who are not available everywhere, and visual inspection alone can be subjective. This paper studies automated skin disease detection and diagnosis using supervised deep learning. Four widely used convolution neural networks, namely ResNet50, VGG16, InceptionV3 and EfficientNetB0, were trained as baselines and compared with a proposed Standard CNN (SCNN) on a seven-class skin disease classification task. All models were evaluated on the same test set of 9,387 images using accuracy, precision, recall, F1-score and confusion matrices. ResNet50, VGG16, InceptionV3 and EfficientNetB0 reached accuracies of 81%, 83%, 87% and 90%, respectively, while the proposed SCNN achieved the highest accuracy of 94%, with macro-averaged precision, recall and F1-score of 0.94. The SCNN also gave the best F1-score in every class, with values between 0.90 and 0.96, and its gains were largest in the classes that the baseline networks confused most often.
The results suggest that adding a self-correction mechanism to a convolutional backbone improves feature learning and generalization compared with standard transfer-learning networks. The study also discusses its limitations, such as the use of a single dataset and the lack of external clinical validation, and outlines directions for explainable and clinically usable skin disease diagnosis systems.