Objective: This study aimed to assess the ability of radiomics features extracted from grayscale B-mode dermatologic ultrasound images to distinguish benign from malignant skin lesions and to compare the performance of multiple machine learning (ML) classifiers trained on different radiomics feature domains, including original, Laplacian-of-Gaussian (LoG), and wavelet-transformed images.
Materials and
Methods: This retrospective study analyzed 190 skin lesions (129 benign, 61 malignant) from a publicly available dermatologic ultrasound dataset. Manual lesion segmentation was performed by an experienced radiologist. Radiomics features were extracted from original grayscale, LoG-filtered, and wavelet-transformed images using PyRadiomics. Low-variance descriptors and highly correlated features were removed, and mutual information–based feature selection was conducted within stratified five-fold cross-validation to avoid information leakage. Logistic regression, support vector machine (RBF), and random forest classifiers were trained using balanced class weights. Performance metrics were computed by pooling predictions from all folds.

Results: A total of 851 radiomics features per lesion were initially extracted and reduced to 143 non-redundant descriptors after preprocessing. Across all models and feature domains, the area under the receiver operating characteristic curve (AUC) ranged from 0.64 to 0.71. The highest performance was achieved by the random forest classifier using the combined feature set (AUC: 0.706; accuracy: 0.716). Wavelet-only features also performed well, with random forest achieving an AUC of 0.705 and the highest accuracy (0.742). Logistic regression applied to original features yielded an AUC of 0.697 and comparatively better sensitivity. Specificity was consistently higher than sensitivity across classifiers.

Conclusion: Radiomics-based analysis of dermatologic ultrasound provides moderate discriminatory ability for benign–malignant classification. Although limited by dataset size and manual segmentation, the findings support radiomics as a quantitative, operator-independent adjunct for ultrasound-based skin lesion evaluation.