Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Two-Stage Training Framework Using Multicontrast MRI Radiomics for IDH Mutation Status Prediction in Glioma

Link

https://dx.doi.org/10.1148/ryai.230218

Indexing

Keywords: Glioma, Isocitrate Dehydrogenase Mutation, IDH Mutation, Radiomics, MRI

Content: MR, NR, OI

RadLex: RID5683

SNOMED: 115240006, 393564001

Author(s)

Nghi C. D. Truong

Chandan Ganesh Bangalore Yogananda

Benjamin C. Wagner

James M. Holcomb

Divya Reddy

Niloufar Saadat

Kimmo J. Hatanpaa

Toral R. Patel

Baowei Fei

Matthew D. Lee

Rajan Jain

Richard J. Bruce

Marco C. Pinho

Ananth J. Madhuranthakam

Joseph A. Maldjian

Organization(s)

The University of Texas Southwestern Medical Center

The University of Texas at Dallas, Department of Bioengineering

New York University Grossman School of Medicine, Departments of Radiology and Neurosurgery

University of Wisconsin–Madison, Department of Radiology

Version

1.0

License

Text: CC BY 4.0

URL: https://creativecommons.org/licenses/by/4.0/

Contact

ude.nretsewhtuostu@gnourT.ihgN

Funding

A.J.M. and J.A.M. supported by NIH/NCI U01CA207091; J.A.M. supported by R01CA260705; R.J.B. supported by NIH 1R01LM013151-01A1.

Ethical review

Institutional Review Board approval was obtained at UTSW with a waiver of consent for the use of retrospective data or public datasets; study was HIPAA compliant.

Date

Published: 2024-05-22

Created: 2023-06-26

References

[1] Truong NCD, Bangalore Yogananda CG, Wagner BC, Holcomb JM, Reddy D, Saadat N, Hatanpaa KJ, Patel TR, Fei B, Lee MD, Jain R, Bruce RJ, Pinho MC, Madhuranthakam AJ, Maldjian JA. "Two-Stage Training Framework Using Multicontrast MRI Radiomics for IDH Mutation Status Prediction in Glioma". Radiology: Artificial Intelligence. 2024 Jul;6(4):e230218.. 2024-05-22. doi:10.1148/ryai.230218. PMID: 38775670. PMCID: PMC11294953.

Model

Architecture

Radiomics-based machine learning with Boruta feature selection and ensemble classifiers (Random Forest and XGBoost) trained via a two-stage balanced-subset framework with probability ensembling.

Availability

Not stated.

Clinical benefit

Noninvasive preoperative prediction of IDH mutation status in glioma, a key prognostic marker that can inform treatment selection and risk stratification.

Clinical workflow phase

Preoperative assessment and clinical decision support.

Degree of automation

Automated pipeline for preprocessing and tumor segmentation (FeTS or other tools), automated radiomics extraction (PyRadiomics), and automated classification; intended as decision support.

Indications for use

Patients with newly diagnosed glioma with available preoperative multicontrast brain MRI (T1w, postcontrast T1w, T2w, T2-FLAIR) and tumor segmentation; prediction of IDH mutation status.

Input

Preoperative multicontrast brain MRI (T1-weighted, postcontrast T1-weighted, T2-weighted, T2-FLAIR) and tumor segmentation masks for whole tumor and/or combined NET+NCR+ED regions; 1197 radiomic features per ROI per sequence extracted with PyRadiomics.

Instructions

Preprocess MRI to 1×1×1 mm3, bias-correct, skull-strip, and coregister; generate tumor masks (WT and NET+NCR+ED where available); extract radiomics features (original and filtered images) with PyRadiomics; apply Boruta feature selection on multiple balanced subsets; train Random Forest or XGBoost classifiers on each subset; ensemble models by averaging predicted probabilities to obtain final IDH mutation prediction.

Limitations

Requires accurate tumor segmentation; variability across datasets in preprocessing and labeling of subregions; training data are imbalanced (overall ~27% IDH-mutated) and addressed via two-stage balanced-subset training; EGD dataset provided only WT masks; no statistical significance testing; performance may depend on MRI acquisition and segmentation quality.

Output

CDEs: RDE748, RDE748.2

Description: Binary classification: predicted IDH mutation status (mutated vs wild type) with associated probability.

Recommendation

Features from combined ROIs (WT and NET+NCR+ED) and the two-stage balanced-subset training/ensembling strategy yielded the most robust performance across external cohorts.

Reproducibility

Model parameters for Boruta and classifiers provided in Supplemental Table S5; PyRadiomics settings and feature lists provided in Supplemental Tables S3–S4; evaluated across multiple independent external datasets with varying acquisition and segmentation approaches.

Use

Intended: Decision support