Abstract

This study presents a comparative analysis of three hybrid classification models—Decision Tree-CNN, Decision Tree-HOG, and Decision Tree-Logistic Regression—for medical image classification, specifically targeting cyst detection. Each model integrates a decision tree classifier with a distinct feature extraction technique to leverage both interpretability and predictive performance. The Decision Tree-CNN model demonstrates strong generalization and high accuracy (98%) by capturing hierarchical image features, while the Decision Tree-HOG model offers computational efficiency but suffers from lower accuracy (81.3%) due to limited feature representation. The Decision Tree-Logistic Regression model outperforms the others, achieving 96.3% accuracy and exhibiting robustness against class imbalance. Extensive evaluations using ROC curves, precision-recall metrics, and PCA visualizations confirm the effectiveness of the proposed hybrids. The findings highlight the potential of combining classical and deep learning methods to develop interpretable and high-performing diagnostic tools for clinical applications. Future work will explore ensemble strategies and advanced data balancing techniques further to enhance classification reliability in real-world medical imaging scenarios.

Key words: Medical Imaging, Hybrid Machine Learning Models, Cyst Detection, Neuroinformatics