Related Experiment Video
Updated: Nov 7, 2025
09:43
Primary Orthotopic Glioma Xenografts Recapitulate Infiltrative Growth and Isocitrate Dehydrogenase I Mutation
Published on: January 14, 2014
8.2K
Deep Learning Can Differentiate IDH-Mutant from IDH-Wild GBM.
Luca Pasquini1,2, Antonio Napolitano3, Emanuela Tagliente3
1Neuroradiology Unit, NESMOS Department, Sant'Andrea Hospital, La Sapienza University, Via di Grottarossa 1035, 00189 Rome, Italy.
Journal of Personalized Medicine
|April 30, 2021
Summary
A new deep learning model accurately predicts isocitrate dehydrogenase (IDH) mutations in glioblastoma multiforme (GBM) using MRI scans. The model achieved 83% accuracy on rCBV perfusion images, aiding in GBM diagnosis.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Oncology
Background:
- Distinguishing isocitrate dehydrogenase (IDH) mutant from wildtype glioblastoma multiforme (GBM) is challenging due to overlapping magnetic resonance imaging (MRI) features.
- While deep learning shows promise for IDH identification in mixed glioma populations, a GBM-specific model is needed.
Purpose of the Study:
- To develop and evaluate a glioblastoma multiforme (GBM)-specific deep learning model for predicting IDH mutation status using multiparametric MRI.
- To assess the performance of convolutional neural networks (CNNs) applied to various MRI sequences for this task.
Main Methods:
- A 4-block 2D CNN model was developed and applied to multiparametric MRI sequences (MPRAGE, T1, T2, FLAIR, rCBV, ADC) from 100 adult patients with WHO grade IV gliomas.
- IDH mutation prediction probability was derived from the softmax activation function's output.
- Model performance was evaluated using accuracy and categorical cross-entropy loss (CCEL) on a test cohort.
Main Results:
- The GBM-specific deep learning model achieved a maximum accuracy of 83% with a CCEL of 0.64 on relative cerebral blood volume (rCBV) maps.
- Other MRI sequences showed varying performance: T1 (77% accuracy), FLAIR (77%), T2 (67%), and MPRAGE (66%).
- ADC maps yielded lower performance compared to other sequences.
Conclusions:
- A GBM-tailored deep learning model can effectively predict IDH mutation status using multiparametric MRI.
- Relative cerebral blood volume (rCBV) perfusion imaging demonstrated the highest predictive accuracy, potentially linked to IDH mutation's effect on neoangiogenesis via hypoxia-inducible factor.
