SKIN LESION CLASSIFICATION FOR MOBILE HEALTH APPLICATIONS: A COMPARATIVE DEEP LEARNING APPROACH

Abstract

The global prevalence of dermatological diseases justifies the need for novel solutions towards early and accessible detection. This study proposes the creation and comparative evaluation of two deep learning models for skin lesion image classification into 14 categories, such as malignant or benign lesions. One was a specially designed Convolutional Neural Network (CNN) trained from scratch, and another used transfer learning on pre-trained MobileNetV2 architecture. The models were trained on a 10,940 image dataset, validated on 1,177 images, and tested on a test set of 1,185 images. The user-defined CNN model had a general test accuracy of 45.91%, while the MobileNetV2 model overwhelmingly performed better with an accuracy of 57.97%. Performance metrics like precision, recall, and F1-score were examined per class and identified particular challenges with class imbalance and visual similarity between some pathologies. The higher accuracy and built-in computational efficiency of the MobileNetV2 model give it favourable candidacy for deployment within the resource-limited mobile setting. This research effectively sets a groundwork automated classification system that can facilitate the creation of an easy-touse mobile application to deliver initial skin lesion analysis and encourage early medical consultation.