Noninvasive Glioma Grading with Deep Learning: A Pilot Study
Summary
This pilot study asks if a deep learning model can assign a WHO 2007 grade (I to IV) to a glioma from preoperative MRI alone. The data come from the PACS archive of the Burdenko National Medical Research Center for Neurosurgery in Moscow.
Data and method
- The archive held 1,280 patients with glial tumors, a verified WHO grade and preoperative MRI from 2009 to 2018. The study kept only contrast-enhanced axial T1 series, because most patients had them. This gave 707 MRI studies: 189, 133, 127 and 258 for grades I, II, III and IV.
- 3D classification: one prediction per patient. Each series becomes one volume of 32 slices at 512 x 512. The model is a DenseNet adapted to 3D input.
- 2D classification: one prediction per slice, over 17,730 slices. The model is ResNeSt200e with ImageNet normalization and light augmentation.
- The split is 80% train, 10% validation, 10% test, with class balance kept. Training uses Adam at learning rate 1e-4 with cross-entropy loss.
Results
| Task and metric | 3D (DenseNet) | 2D (ResNeSt200e) |
|---|---|---|
| 4 grades, accuracy | 83% | 50% |
| 4 grades, ROC AUC (mean one-vs-rest) | 0.95 | 0.72 |
| Low vs high grade, accuracy | 67% | 61% |
| Low vs high grade, ROC AUC | 0.76 | 0.73 |
| Low vs high grade, sensitivity | 58% | 44% |
| Low vs high grade, specificity | 78% | 81% |
The 3D model was strongest on grades I and IV (F1 0.88 and 0.90) and weaker on grades II and III (F1 0.76 and 0.67). The 2D model failed on grades II and III (F1 0.17 and 0.04). The authors give one likely cause: no slice was labeled by hand, so slices with no tumor still carried the patient's grade.
Limits
The authors call this a pilot. The sample is small, and slice count, resolution and scanner differ between patients. Only one MRI modality is used. The result shows that contrast-enhanced T1 images separate WHO grades. It is not a validated clinical tool. The same 707-study cohort and the 83% / 0.95 result also appear in the journal paper MR-guided non-invasive typing of brain gliomas using machine learning.
Cite as
@inproceedings{danilov2022noninvasive,
title={Noninvasive Glioma Grading with Deep Learning: A Pilot Study},
author={Danilov, Gleb and Korolev, Vladislav and Shifrin, Michael and Ilyushin, Eugene and Maloyan, Narek and Saada, Daniel and Ishankulov, Timur and Afandiev, Ramin and Shevchenko, Alexander and Konakova, Tatyana and Tsukanova, Tatyana and Shugay, Svetlana and Pronin, Igor and Potapov, Alexander},
booktitle={MEDINFO 2021: One World, One Health},
series={Studies in Health Technology and Informatics},
volume={290},
pages={675--678},
publisher={IOS Press},
year={2022},
doi={10.3233/SHTI220163}
}