Author: Rituparna Basak, Maede Boroji, Renee F Cattell, Vahid Danesh, Imin Kao, Kartik Mani, Xin Qian, Samuel Ryu, Tiezhi Zhang 👨🔬
Affiliation: Stony Brook Medicine, Stony Brook University, Washington University in St. Louis, Stony Brook University Hospital 🌍
Purpose: Fundamental qualitative characteristics physicians use to differentiate skin lesion subtypes include asymmetry, border irregularity, and color. Radiomic features have potential to quantify these features. This study aims to (1) develop a classification model for skin lesions subtypes using radiomic features and (2) evaluate the added value of incorporating the border radiomic features in the classification process.
Methods: 10,015 skin lesion images from the HAM 10,000 dataset were analyzed. Regions of interest were visible skin lesions and a ring +/- 50 pixels of visible lesion to capture border features. PyRadiomics was used to extract features including first-order statistics, Gray Level Co-occurrence Matrix, and Gray Level Size Zone Matrix. Sample sizes of training and testing sets were 8,015 and 2,000, respectively. Principal component analysis was applied to reduce the feature dimension to 35 features. Synthetic Minority Over-sampling Technique was applied to the training set to address class imbalance. Random Forest classifier with 500 decision trees was trained first on lesion only and then on combined lesion-border features. Model performance was assessed using accuracy, sensitivity and specificity.
Results: Using lesion features alone, testing set performance (accuracy, sensitivity, specificity) of average of 5-fold cross-validation for three largest classes are: melanocytic nevi (64%, 36%, 61%), melanoma (44%, 87%, 15%), and benign keratosis-like lesions (60%, 88%, 8%). After incorporating border features, testing performance improved significantly with the following results: melanocytic nevi (99%, 83%, 99%), melanoma (89%, 98%, 62%), and benign keratosis-like lesions (99%, 99%, 99%). The inclusion of border features notably enhanced the model performance for these lesion subtypes.
Conclusion: Radiomic features derived from the skin lesion demonstrate robust predictive capability in classification. Features encompassing the peri-lesional/border area improve model’s performance, thereby improving the accuracy of lesion classification. This aligns with clinical emphasis on border characteristics, demonstrating the potential of radiomics to augment diagnostic precision.