@article{Liu2026, 
author = {Yihao Liu and Zhihao Cui and Liming Li and Junjie You and Xinle Feng and Ruixuan Huang and Jianxin Wang and Xiangyu Wang and Qing Liu and Minghua Wu},
title = {GMMAS: Glioma Multiparametric MRI Analysis System for Fully‐Automated Layered Tumor Diagnosis and Prognostic Evaluation},
year = {2026},
journal = {Medicine Advances},
volume = {4},
number = {2},
pages = {164-178},
keywords = {glioma, layered diagnosis, prognosis evaluation},
url = {https://www.sciopen.com/article/10.1002/med4.70064},
doi = {10.1002/med4.70064},
abstract = {BackgroundGliomas are the most common primary tumors of the central nervous system. Multiparametric magnetic resonance imaging (mpMRI) is widely used for preliminary screening and plays a crucial role in auxiliary diagnosis, therapeutic efficacy, and prognostic evaluation. We aim to develop a fully automated system based on mpMRI for accurate glioma diagnosis and reliable prognostic assessment.MethodsIn this study, we collected a total of 1962 mpMRI samples and used them to develop a Glioma Multiparametric MRI Analysis System (GMMAS). This system leverages an uncertainty‐based semi‐supervised multitask learning architecture and simultaneously outputs a segmentation of the tumor region, the histological subtype of the glioma, the isocitrate dehydrogenase mutation genotype, and the status of 1p/19q chromosome disorder. Moreover, by utilizing a contrastive learning‐based adaptation module for cross‐modal feature extraction, GMMAS exhibits robustness when certain magnetic resonance imaging modalities are absent. Finally, based on the GMMAS analysis outputs, we created a user‐friendly platform for both doctors and patients. By integrating medical knowledge through the retrieval‐augmented generation technique, we introduced GMMAS‐generative pre‐trained transformer to generate personalized prognostic evaluations and offer treatment suggestions for glioma patients.ResultsIn the tumor segmentation task, the Dice values for the whole tumor, tumor core, and edema region reached 0.940 ± 0.037, 0.919 ± 0.092, and 0.870 ± 0.101, respectively. For glioma subtyping, the internal validation accuracies for differentiating glioblastoma from low grade glioma, predicting IDH mutation, and predicting 1p/19q co‐deletion were 0.941, 0.950, and 0.896, respectively. In the external validation cohorts, the average area under the curve values for these three prediction tasks were 0.905, 0.924, and 0.896, respectively.ConclusionsGMMAS is a fully automated platform that delivers accurate and comprehensive diagnostic predictions and prognostic reports for glioma patients via non‐invasive approaches.}
}