Comparative Analysis of Machine Learning Algorithms for Dermoscopic Skin Lesion Recognition

Abstract

Automated classification of dermoscopic skin lesions can support computer-aided diagnosis, but performance depends on both the classifier and the image representation. This study compares six classical machine-learning algorithms—KNN, SVM-RBF, SVM-Linear, LR, RF, and XGBoost—for six-class recognition using the Derm7pt dataset. After preprocessing and filtering, 1,002 dermoscopic images were retained. The images were consolidated into six diagnostic categories: BCC, MEL, MISC, NEV, SK, and VASC, and represented as 32 × 32 RGB images converted into 3,072-dimensional raw-pixel vectors. The official Derm7pt partition was maintained, with 609 images in the development set and 393 images in the independent test set. Classifiers were evaluated using accuracy, macro-precision, macro-recall, macro-F1, macro-AUC, macro-average precision, and log-loss. Baseline XGBoost achieved an accuracy of 0.6132, a macro-F1 of 0.3062, and a macro-AUC of 0.7879. Hyperparameter optimisation was performed for all six classifiers using stratified three-fold cross-validation with macro-F1 as the model-selection criterion. The effects of tuning varied among classifiers, indicating that classifier-level optimisation alone does not overcome the limitations of the low-resolution raw-pixel representation. Class-level analysis also showed uneven recognition across diagnostic categories, particularly for minority classes. The findings support the use of multiple evaluation metrics and motivate investigation of stronger feature representations and class-balancing strategies.