A comparative study of hybrid decision tree–deep learning models in the detection of intracranial arachnoid cysts

Intracranial arachnoid cysts are fluid-filled lesions within the arachnoid membrane, which pose significant diagnostic challenges due to their varying sizes, subtle radiographic characteristics, and often unclear clinical correlations. Traditional diagnostic methods, such as MRI or CT imaging, rely on expert interpretation but suffer from issues like inter-observer variability and diagnostic delays, especially for small or atypically located cysts. To address these challenges, this study integrates machine learning (ML) and deep learning (DL) techniques into neuroimaging diagnostics, introducing three novel hybrid models: DecisionTree-ViT, DecisionTree-Random Forest, and DecisionTree-ResNet50. The DecisionTree-Random Forest hybrid model showed remarkable performance, achieving 96.3% accuracy and 0.98 AUC in differentiating arachnoid cysts from normal cerebrospinal fluid spaces and other intracranial cystic lesions. This model combines deep learning's pattern recognition strengths with decision tree transparency, meeting the clinical need for both accuracy and explainability. The DecisionTree-ResNet50 variant excelled in detecting small (<1 cm) cysts, with a sensitivity of 89.7%, outperforming standalone ResNet50 (82.4%). Specialized contrast-enhancement protocols and anatomically constrained augmentation techniques were applied to address class imbalance and improve model calibration. The DecisionTree-ViT model also demonstrated strong performance, with 94% accuracy and well-calibrated confidence estimates, making it reliable for clinical decision-making. The study compares these hybrid models against pure deep learning and traditional machine learning approaches, highlighting their superior performance in challenging diagnostic scenarios. The integrated interpretability features allow radiologists to validate algorithmic findings, fostering trust in AI-assisted diagnostics. This research showcases the potential of hybrid AI models to transform neuroimaging diagnostics and improve patient outcomes.