Brain Tumor Segmentation and Grading on MRI Using Deep Learning: A Systematic Literature Review and Benchmark-Driven Comparative Analysis
TL;DR
A benchmark-driven comparative synthesis ofMRI segmentation and histological grading in brain tumor segmentation and histological grading and an adapted QUADAS framework are conducted, finding external generalizability, uncertainty quantification, reproducibility, and prospective validation remain weak.
Abstract
Background: Magnetic resonance imaging (MRI) is central to brain tumor segmentation and histological grading, and deep learning (DL) has transformed both tasks. Existing reviews rarely span the 2017–2026 architectural arc from CNNs and U-Net variants to transformers and foundation models or appraise reproducibility and clinical-translation readiness. Methods: This preregistered systematic review (PRISMA 2020, PRISMA-S) searched Google Scholar, PubMed/MEDLINE, and IEEE Xplore on 21 May 2026 (January 2017–May 2026). A single reviewer performed screening, extraction and QUADAS-AI appraisal with repeated checks on separate days; therefore, the synthesis is presented as a transparent descriptive review rather than a pooled meta-analysis. Records were screened against a priori eligibility criteria; primary experimental studies entered the synthesis and review articles formed a contextual corpus. Methodological quality was appraised using an adapted QUADAS framework (“QUADAS-AI”). Heterogeneity precluded statistical pooling, so a benchmark-driven comparative synthesis was conducted. Results: The search retrieved 33,982 records (Google Scholar 23,400; PubMed/MEDLINE 5349; IEEE Xplore 5233). After deduplication and screening, 141 full texts were assessed; one report published before the eligibility window was excluded, leaving 140 included studies: 117 primary (39 contributed to the BraTS Dice benchmark sub-set) and 23 contextual reviews. U-Net variants (42%) and hybrid CNN–transformer architectures (40%) dominate, followed by CNN classifiers (14%) and vision transformers (3%). On BraTS 2021, nnU-Net and Swin UNETR reach DSC 0.93/0.90/0.85 (whole tumor/core/enhancing); reported external evaluations show 2.5–15 percentage-point performance drops under domain shift. Conclusions: External generalizability, uncertainty quantification, reproducibility, and prospective validation remain weak; ten research priorities are proposed. Registration: OSF, DOI 10.17605/OSF.IO/C2QA6; registered 20 May 2026.