ResNet50-Based Multi-Class Skin Lesion Classification with Image Suitability Screening in a Web-Based Decision-Support System

Abstract

Limited access to specialist dermatology and the variability of image-based assessment motivate the use of artificial intelligence as a preliminary decision-support tool. This study evaluated five convolutional neural network architectures—MobileNetV2, EfficientNetB0, ResNet50, DenseNet121 and InceptionV3—for seven-class skin-lesion classification using the HAM10000 dataset, and integrated the selected classifier into a web-based prototype with image-suitability screening. The evaluation used test accuracy, precision, recall, F1-score and training time. ResNet50 produced the highest initial test accuracy (72.06%) and F1-score (74.00%), although MobileNetV2 and EfficientNetB0 reported slightly higher precision. ResNet50 was therefore selected and retrained for a further 30 epochs. Its test accuracy increased to 80.24% and its F1-score to 80.85%, corresponding to gains of 8.18 and 6.85 percentage points, respectively. The implemented inference workflow converts uploaded images to RGB, resizes them to 224 × 224 pixels, normalises them, screens their suitability through a Gatekeeper stage, and then applies the ResNet50 classifier when an image is accepted. The web application further supports secure login, email one-time-password verification, role-based access control, patient records, diagnostic history and report generation. The study demonstrates comparative model selection and end-to-end system integration, but it does not establish clinical effectiveness. No external dataset, locally collected Somali clinical images, dermatologist comparison or prospective patient evaluation was used. The system should therefore be interpreted as a research-stage decision-support prototype requiring external and clinical validation before use in patient care.