Work

Explainable AI for MRI Brain Tumor Classification: A Scoping Review of Clinical Validation

Year: 2026
Type: preprint
Study design: Systematic Review
Abstract: Machine learning and deep learning models for magnetic resonance imaging (MRI)-based brain tumor classification now report accuracies above 98% on public benchmarks, yet clinical use has not followed. Explainable artificial intelligence (XAI) is presented as the bridge, but whether the explanations are correct is seldom tested. This scoping review, following PRISMA-ScR-compatible reporting conventions, searched PubMed, IEEE Xplore, SpringerLink, ScienceDirect and Google Scholar on 13 July 2026 for studies published from 2020 to 2026. Of 105 screened records, 50 primary studies and 12 reviews were included. The 22 primary studies extracted at full text were audited with a five-level explanation-validation hierarchy and a nine-item reporting-integrity instrument, organised around five research questions. Twelve of the 22 studies applied no explanation; seven assessed explanations by visual narration, one against a segmentation mask and two with internal faithfulness metrics. None validated explanations against radiologist annotation (0/22; 95% CI 0–15.4%). Median best accuracies were similar for handcrafted (98.5%), hybrid (98.7%) and end-to-end deep pipelines (98.1%). Only two studies reported a verifiable patient-level split, three validated externally, two released code, two reported calibration, and four reported headline figures inconsistent with their own data. Interpretability is better described by where it sits in the pipeline, and whether it has been checked, than by the modelling paradigm. We summarise tumor types, imaging modalities, classification and segmentation methods and public datasets, and propose a 12-item checklist that makes agreement with expert annotation, patient-level partitioning and deployment cost routine reporting items.
Language: English
Open Access status: green
OpenAlex ID: W7226029485
Cites:
Cited by:
Related to: 10