Automated Dermatological Diagnosis: A Convolutional Neural Network Framework for Multi-Class Skin Lesion Recognition

Abstract

Dermatological conditions impact a vast majority of the global population, with early identification being critical for improving patient outcomes, particularly in life-threatening cases such as melanoma. Traditional diagnostic workflows rely heavily on expert dermatological evaluation of dermoscopic imagery, a process that is time-intensive, subjective, and geographically constrained. To address these limitations, this study presents an automated, deep learning-driven classification pipeline designed for multi-category skin lesion recognition. Utilizing the publicly available HAM10000 dataset, which encompasses seven distinct lesion categories, the proposed system integrates standardized preprocessing, strategic data augmentation, and a custom-designed Convolutional Neural Network (CNN). The model systematically extracts hierarchical visual representations and maps them to diagnostic classes through a softmax-enabled output layer. Experimental evaluation demonstrates a test accuracy of 97.40%, with strong precision, recall, and F1-score distributions across all categories. The findings confirm that tailored CNN architectures can deliver reliable, rapid, and clinically relevant lesion classification, positioning artificial intelligence as a viable decisionsupport mechanism in dermatological practice.