Skin Lesion Image Classification With Tree-Based Ensembles: Benchmarking Random Forest and Gradient Boosting.

Area of Science:

• Dermatology
• Medical Imaging
• Machine Learning

Background:

• Skin cancer diagnosis relies on visual assessment, facing standardization and accessibility issues.
• Deep learning models show promise but require significant resources and lack interpretability.
• This study explores tree-based ensemble methods as a viable alternative.

Purpose of the Study:

• To evaluate the accuracy of Random Forest (RF) and Gradient Boosted Decision Trees (GBDT) for classifying four common skin lesions.
• To compare the performance of these tree-based models against a deep learning benchmark.
• To assess the interpretability and training efficiency of the proposed methods.

Main Methods:

• Utilized 8,000 dermoscopic images across four lesion classes (BCC, BKL, MN, melanoma).
• Applied image preprocessing, handcrafted feature extraction (Haralick, LBP, RGB histograms), and feature normalization.
• Optimized RF and GBDT hyperparameters using Bayesian search and cross-validation, benchmarking against MobileNetV2.

Main Results:

• GBDT achieved 89% accuracy and 0.88 F-score; RF achieved 86% accuracy and 0.85 F-score.
• Both ensembles exceeded 0.94 area under the ROC curve for melanoma detection.
• Tree-based models trained over 10 times faster than the deep learning benchmark, with SHAP analysis revealing interpretable feature contributions.

Conclusions:

• Traditional machine learning, specifically tree-based ensembles, offers an effective and interpretable approach for skin lesion classification.
• These models provide a practical alternative to deep learning, especially in resource-constrained environments.
• Feature engineering combined with tree-based methods can achieve performance comparable to deep learning with enhanced transparency.