SKIN DISEASE DETECTION USING DEEP LEARNING WITH CNN AND ResNet50

The exponential rise in skin cancer incidence globally has underscored the urgent need for accurate, automated, and
accessible diagnostic tools. Traditional dermatological diagnosis — reliant on visual inspection, dermoscopy, and
histopathological biopsy — is inherently subjective, resource-intensive, and inaccessible in underserved settings. This paper presents a deep learning-based Skin Disease Detection system integrating Convolutional Neural
Networks (CNN) with the ResNet50 architecture to enable precise, automated classification of skin lesions. The
proposed system leverages transfer learning and fine-tuning on large annotated dermatoscopic datasets, including
the HAM10000 benchmark, to distinguish between benign and malignant lesion types across seven classes: AKIEC,
BCC, BKL, DF, MEL, NV, and VASC. The pipeline encompasses image preprocessing, data augmentation, feature
extraction via ResNet50, and softmax-based classification through a fully connected layer. Experimental evaluation
demonstrates an overall classification accuracy exceeding 90%, with high sensitivity and specificity, and an AUCROC score confirming the model's discriminative power. The system is implemented using TensorFlow and MATLAB, with a user-friendly graphical interface enabling healthcare practitioners to upload dermoscopic images and receive real-time diagnostic predictions. Comparative analysis confirms that the proposed CNN-ResNet50 system outperforms traditional image-processing-based classifiers and selected deep learning baselines on identical datasets. This work demonstrates the feasibility of an accurate, scalable, and clinically viable automated skin cancer classification system for deployment in diverse healthcare environments.