Machine Learning-Based Skin Disease Diagnosis Using Dermoscopic Images: A Multi-Class Classification Framework

Abstract

Accurate identification of skin lesions from dermoscopic images is
an important task in computer-assisted medical diagnosis. This study
presents a machine-learning framework for classifying dermoscopic
images into multiple skin-lesion categories using image preprocessing,
feature extraction, feature selection, and supervised classification. The
HAM10000 dataset is considered for experimental evaluation because
it provides 10,015 dermoscopic images covering seven commonly
investigated diagnostic categories: melanoma, melanocytic nevus,
basal cell carcinoma, actinic keratosis/Bowen disease, benign keratosis,
dermatofibroma, and vascular lesions. The proposed approach does
not rely on deep neural networks. Instead, descriptive characteristics
related to lesion colour, texture, shape, and intensity are extracted from
the images and supplied to conventional machine-learning classifiers.
Support Vector Machine, Random Forest, and k-Nearest Neighbour
classifiers can be evaluated under the same experimental conditions.
Because the dataset contains a considerable difference in the number of
samples available for individual classes, class-wise sensitivity, specificity,
precision, recall, F1-score, balanced accuracy, and one-vs-rest AUC are
considered in addition to overall accuracy. A seven-class confusion
matrix is also used to examine errors between individual diagnostic
categories. Training and testing curves, where applicable, can be
used to monitor classifier behaviour, while prediction time provides
an indication of computational efficiency. The framework is intended
to provide a transparent and computationally manageable approach
for automated skin-lesion classification. The study also discusses the
limitations associated with class imbalance, image acquisition conditions,
dataset diversity, and generalization to images obtained from clinical
environments.