SKIN CANCER CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Early assessment of suspicious skin lesions can improve access to care, but dermoscopic interpretation depends on specialist expertise and automated classifiers must be evaluated carefully under severe class imbalance. This paper presents a seven-class convolutional neural network pipeline and a cloud-connected desktop prototype developed from a thesis-scale study. The classifier uses an ImageNet-initialized MobileNetV2 backbone followed by global average pooling, dropout, a 128-unit ReLU layer, and a seven-way softmax output. The HAM10000 dataset contains 10,015 dermoscopic images and was stratified into training, validation, and test subsets of 7,010, 1,502, and 1,503 images. Images were resized to 224 × 224 pixels, normalized, augmented, and trained with class weighting in TensorFlow/Keras on Google Colab. On the held-out test set, the reported accuracy was 93.20%, with weighted precision, recall, and F1-score of 94.06%, 93.21%, and 93.51%, respectively. Melanoma recall was 88.62%, whereas dermatofibroma precision was 45.16% with only 17 test examples, underscoring the effect of class imbalance. A Flask REST API and PySide6 client produced an average end-to-end response time of 3.3 s. The results support the feasibility of the prototype as a decision-support tool, while the absence of clinical and external validation limits claims of clinical readiness