Discrimination of glioma from brain metastasis can pose a diagnostic challenge in management of patients. Several classical ML and Deep Learning (DL) models have been proposed to address this and achieve accurate classification. While studies consistently cite good accuracy, it remains unclear which algorithm should be incorporated into clinical practice.

Our study identified 25 eligible studies. The most predominant source of data were single-centre hospital datasets (n=23). Only one study reported validation on external datasets. On average, datasets included 121.0 ± 91.88 patient studies. For internal validation, cross validation was most frequently performed. Among the best classifiers reported in every study, with classification accuracy altogether averaging at 0.90 ± 0.09, SVM were most frequent (n=9), followed by miscellaneous DL algorithms (n=4). Average TRIPOD adherence score was 0.45 ± 0.12. Poorest performance was detected in Title, Abstract, and Missing Data sections with TRIPOD adherence indices of 0, 0 and 0.13, respectively.

Our findings show lack of reproducibility in ML studies on distinction between glioma and brain metastasis, mainly due to small datasets and lack of model validation. In consequence, we recommend multi-centre collaboration, external validation, and adherence to risk of bias assessment tools.