Deep learning for brain tumour diagnosis on MRI: a systematic review of reported performance, evaluation design and reporting completeness across classification, detection and segmentation studies (2021–2026)
Year: 2026
Type: preprint
Study design: Systematic Review
Abstract: This systematic review examines deep learning models applied to brain MRI for tumour classification, detection and segmentation, in studies published from 2021 onward. Its purpose is not to catalogue architectures or rank reported accuracies, but to establish the conditions under which those accuracies were obtained. Existing reviews of this field report performance figures clustered between 97% and 100% without extracting the information needed to interpret them: how data were partitioned, whether images from the same patient appear on both sides of the train/test boundary, and whether augmentation preceded or followed splitting. Where many studies evaluate on the same public collection, convergent high accuracy has been read as convergent evidence, although it is not independent replication. No prior review of this field applies a formal quality appraisal, and none distinguishes the task a study claims to perform from the task its reported metrics actually evidence. The review therefore asks five questions: how the literature is distributed across task, architecture, clinical target and data regime, and how many distinct evaluation datasets it rests on; how completely studies report split protocol, patient-level separation, augmentation order, class balance, evaluation unit and variability; how reported performance differs when studies are stratified by split protocol and dataset, and which findings survive excluding inadequately reported studies; how often claimed task and evidenced task diverge; and what evidence of generalisation the corpus contains.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7202045910