MR-guided Non-invasive Typing of Brain Gliomas Using Machine Learning
Summary
This journal paper tests deep learning for glioma typing by the 2007 WHO grade (I to IV) from preoperative MRI. Here "typing" means the grade. The paper does not predict molecular markers such as IDH mutation or 1p/19q codeletion.
Data and method
- The data are MRI scans of patients with glial tumors treated at the Burdenko National Medical Research Center for Neurosurgery. All patients had contrast-enhanced MRI before surgery.
- Two models: ResNeSt200e on 2D slices and DenseNet on 3D volumes. Both classify the tumor into 4 classes, WHO grades I to IV.
- Training uses a random 80% of the studies. The other 20% are the validation and test samples.
Results
The analysis included 707 contrast-enhanced T1-weighted studies. The 3D DenseNet model gave the best prediction of WHO grade: accuracy 83%, AUC 0.95. The authors note that other groups report similar results with other methods. Their conclusion is that axial contrast-enhanced T1 images can be graded by 2007 WHO class with deep learning models.
This page uses the published abstract. The cohort and the headline numbers match the conference pilot Noninvasive Glioma Grading with Deep Learning, which also gives the per-grade and binary results.
Cite as
@article{danilov2022mr,
title={{MR}-guided non-invasive typing of brain gliomas using machine learning},
author={Danilov, G. V. and Pronin, I. N. and Korolev, V. V. and Maloyan, N. G. and Ilyushin, E. A. and Shifrin, M. A. and Afandiev, R. M. and Shevchenko, A. M. and Konakova, T. A. and Shugai, S. V. and Potapov, A. A.},
journal={Zhurnal Voprosy Neirokhirurgii imeni N. N. Burdenko},
volume={86},
number={6},
pages={36--42},
year={2022},
doi={10.17116/neiro20228606136}
}