Deep learning for prediction of isocitrate dehydrogenase mutation in gliomas: a critical approach, systematic review and meta-analysis of the diagnostic test performance using a Bayesian approach.

BACKGROUND

Conventionally, identifying isocitrate dehydrogenase () mutation in gliomas is based on histopathological analysis of tissue specimens acquired via stereotactic biopsy or definitive resection. Accurate pre-treatment prediction of mutation status using magnetic resonance imaging (MRI) can guide clinical decision-making. We aim to evaluate the diagnostic performance of deep learning (DL) to determine mutation status in gliomas.

METHODS

A systematic search of Cochrane Library, Web of Science, Medline, and Scopus was conducted to identify relevant publications until August 1, 2021. Articles were included if all the following criteria were met: (I) patients with histopathologically confirmed World Health Organization (WHO) grade II, III, or IV gliomas; (II) histopathological examination with the mutation; (III) DL was used to predict the mutation status; (IV) sufficient data for reconstruction of confusion matrices in terms of the diagnostic performance of the DL algorithms; and (V) original research articles. Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) and Checklist for Artificial Intelligence in Medical Imaging (CLAIM) was used to assess the studies' quality. Bayes theorem was utilized to calculate the posttest probability.

RESULTS

Four studies with a total of 1,295 patients were included. In the training set, the pooled sensitivity, specificity, and area under the summary receiver operating characteristic (SROC) curve were 93.9%, 90.9% and 0.958, respectively. In the validation set, the pooled sensitivity, specificity, and area under the SROC curve were 90.8%, 85.5% and 0.939, respectively. With a known pretest probability of 80.2%, the Bayes theorem yielded a posttest probability of 97.6% and 96.0% for a positive test and 27.0% and 30.6% for a negative test for training sets and validation sets, respectively.

DISCUSSION

This is the first meta-analysis that summarizes the diagnostic performance of DL in predicting mutation status in gliomas via the Bayes theorem. DL algorithms demonstrate excellent diagnostic performance in predicting mutation in gliomas. Radiomic features associated with mutation, and its underlying pathophysiology extracted from advanced MRI may improve prediction probability. However, more studies are required to optimize and increase its reliability. Limitations include obtaining some data via email and lack of training and test sets statistics.