Segmentation-Free Preoperative 3D MRI Classification of Low-Grade Versus High-Grade Glioma Using Task-Oriented Neural Architecture Search

Description

Christos Ch. Andrianos and co-authors grade brain gliomas from MRI scans without first outlining the tumor, in a 2026 methods study. Their network was itself designed by an automated architecture search over a three dimensional backbone, then strengthened with residual links and attention blocks. Training and validation used preoperative scans from 1,194 patients across four public cohorts, with a fifth collection held out for external testing.

The model separated low grade from high grade tumors with about eighty-eight percent accuracy internally, dropping to roughly seventy-five percent on the unseen cohort, still ahead of benchmark networks. Heat map analysis showed it looking at the tumor region despite never being taught outlines, checked against available hand drawn masks. The same machinery predicted a key mutation with about ninety percent accuracy and handled multi grade sorting respectably. The authors pitch segmentation free grading as a lighter path to computer aided diagnosis.

Catalog record

Source: https://europepmc.org/articles/PMC13301297