Machine learning methods in the detection of melanocytic lesions
Loading...
Date
2026-02-04
Authors
Ποσονίδης, Αθανάσιος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Melanoma is one of the deadliest cancers. Accurate and fast detection can influence drastically the outcome of the patient that is diagnosed.
The aim of this study is to provide a fast and accurate method based on classical machine learning methods, due to their explainability and lack of black box architecture. The dataset was created by Kalloniati Evaggelia during her PhD, implementing an SVM classifier, a Random Forest and Logistic Regression. In her study, the comparison of SVM, logistic regression and random forests showed that Random forests provided the best accuracy, followed by Logistic regression, with the worst accuracy showed by SVM.
In this study, a Naïve Bayes classifier was applied in the same dataset, which was split into subsets of Low Risk-Medium Risk, Low Risk-High Risk and Medium Risk-High Risk cases. Due to the small dataset size, 5-fold and 10-fold Cross-Validation was used. The results showed that differences in Cross-Validation were minimal, with the classifier exhibiting highest performance (AUC=1) for Low-High Risk classification with morphological features. For Low-Medium Risk classification, features relevant to texture heterogeneity, gray-level non-uniformity and directional intensity variations were able to perform better in the task of finer distinction. For Medium-High Risk classification, features that carry information about morphology and tissue irregularities and complexity were among the ones that could effectively differentiate the lesions.
Also, a Minimum Redundancy – Maximum Relevance algorithm was used for feature selection, which provided three features on each subset that had the highest relevance for class differentiation. These features performed well (AUC>0.81) on the Naïve Bayes Classifier. These features were combined and used with the Naïve Bayes Classifier with 5-fold Cross-Validation and provided a significant improvement, with an AUC>0.91 in all datasets.
For assessing feature importance, a Random Forest Feature Importance algorithm was used, measuring SHAP, MDI and Permutation importance. Features that were exhibited higher SHAP and MDI values were morphological, texture and intensity heterogeneity based. Differences were exhibited when features that performed well in the previous algorithms, had reduced importance in the Random Forest due to correlations. While SHAP and MDI could be interpreted, Permutation was unreliable in some datasets, due to the small dataset size.
These results demonstrate that classical machine learning can effectively classify skin lesions, even with small, imbalanced datasets, as shown by mRMR and RFFI algorithms assessment and can be used for future dermatological applications.
Description
Keywords
Machine learning, Melanoma detection, Naive bayes, Random forest, Feature selection, Feature importance