Abstract— This paper presents a comparative analysis of three hybrid Decision Tree–based models—DT-CNN, DT-HOG, and DT-LR—designed for medical image classification. The aim is to evaluate how different feature extraction strategies influence the performance and generalization ability of the Decision Tree classifier. The DT-CNN model achieved the highest validation accuracy (98%) due to its deep feature representation capability, while the DT-LR model obtained the highest test-set accuracy (96.3%), demonstrating superior generalization performance. In contrast, the DT-HOG model achieved an accuracy of 81.3%, reflecting the limitations of handcrafted features for subtle lesion characterization. The results indicate that although CNN-based features provide strong representation power, the Logistic Regression–Decision Tree hybrid exhibits more stable and reliable performance on unseen data. These findings highlight the importance of selecting appropriate feature extractors when designing hybrid machine-learning pipelines for medical image classification.