The literature search yielded 11,727 studies and 1,141 underwent full-text review. Data was extracted from 23 publications. Nineteen studies used cML only, two DL only, and two both. Analyzed databases had an average size of 99 patients/study. 26.1% of publications reported external validation. The most tested ML and deep learning algorithms were Support Vector Machines (SVM) and Multilayer Perceptron Networks (MLP), respectively. For cML a Logistical Regression model achieved the highest AUC (0.961) in external validation, while an MLP achieved the highest for DL (0.947).  Both were trained on conventional radiomic features from routine and DWI sequences. End-to-end classifiers like Convolutional Neural Networks (CNN) achieved lower AUC in external validation (0.49 and 0.89).

AI-based methods for differentiating gliomas and PCNSL have been reported and show that ML models can achieve AUC > 0.94 in external validation. Classifiers using preextracted handcrafted features performed better than end-to-end deep classifiers. CNNs, when not regularized properly, are prone to overfitting and benefit from large datasets. With few studies using DL algorithms further research into novel DL-based approaches is recommended. Additionally, most studies lack large datasets and external validation, increasing the risk of overfitting.