Glioblastoma and lymphoma are challenging to differentiate due to their similar morphology, making preoperative MRI diagnosis often equivo- cal. Accurate differentiation is crucial due to different treatment pathways, and the use of corticosteroids in lymphoma complicates pathological diag- nosis. Neural networks, particularly convolutional neural networks, show promising results, but their black-box nature raises concerns about inter- pretability and explainability, which is essential for general data protection regulation compliance in clinical settings. While previous studies have in- vestigated deep learning for brain tumor differentiation, interpretability and explainability have not been sufficiently explored. This work presents a deep learning approach for glioblastoma and lymphoma differentiation using histopathological diagnosis as ground truth. The aim is to develop a clinician decision support system and to investigate interpretability and explainability using saliency maps, Monte Carlo dropout and deep ensem- bles. The model achieved an area under the receiving operative character- istics curve of 0.9 and an accuracy of 0.8, outperforming the average radiol- ogist’s accuracy of 0.79. However, challenges remain, including robustness, shifts in data distribution and clinical usability. Future work should ad- dress these issues to facilitate the clinical implementation of deep learning based tumor differentiation.
Figure 1: ROC curve from the original publication of the initial balanced test data set.
Figure 2: Saliency maps of the original publication of the initial balanced test data set.
- Linux (Tested on Ubuntu 22.04)
- NVIDIA GPU (Tested on Nvidia GeForce RTX 2080 Ti x 12 on local workstations)
- Python (3.10), nibabel (5.1.0), torch (2.1.2), monai (1.2.0), numpy (1.24.3), tqdm (4.62.3), scikit-learn (1.3.0), matplotlib (3.4.3).
python main.py --batch_size 1 --path /your/path/to/the/data --seed 42 --model /Path/To/Model --device cuda Please structure your dataset as follows:
path/
├── Glioblastom
├── img_caseG1_0000.nii.gz (FLAIR)
├── img_caseG1_0001.nii.gz (T1)
├── img_caseG1_0002.nii.gz (T1ce)
├── img_caseG1_0003.nii.gz (T2)
├── img_caseG1_seg.nii.gz (segmentation)
├── img_caseG2_0000.nii.gz (FLAIR)
├── img_caseG2_0001.nii.gz (T1)
├── img_caseG2_0002.nii.gz (T1ce)
├── img_caseG2_0003.nii.gz (T2)
├── img_caseG2_seg.nii.gz (segmentation)
└── ...
└── Lymphom
├── img_caseL1_0000.nii.gz (FLAIR)
├── img_caseL1_0001.nii.gz (T1)
├── img_caseL1_0002.nii.gz (T1ce)
├── img_caseL1_0003.nii.gz (T2)
├── img_caseL1_seg.nii.gz (segmentation)
├── img_caseL2_0000.nii.gz (FLAIR)
├── img_caseL2_0001.nii.gz (T1)
├── img_caseL2_0002.nii.gz (T1ce)
├── img_caseL2_0003.nii.gz (T2)
├── img_caseL2_seg.nii.gz (segmentation)
└── ...where G1 or L1 are the identifiers for the subjects. Each subject is expected to have all four modalities available 0000, 0001, 0002, 0003 and a segmentation.
This project is made avialble under the CreativeCommons License. See the license file for more details.
If you find our work useful in your research or use parts of this code please consider citing our paper:
@article{GlioLymphDifferentiation,
title={Development of a deep-learning model for the preoperative diagnosis of Primary Central Nervous System Lymphoma},
author={Paul Vincent Naser, Miriam Cindy Maurer, Maximilian Fischer, Kianush Karimian-Jazi, Chiraz Ben-Salah, Awais Akbar Bajwa, Martin Jakobs, Jessica Jesser, Martin Bendszus, Klaus Maier-Hein, Sandro M. Krieg, Peter Neher, Jan-Oliver Neumann},
journal={iScience},
year={2024},
publisher={iScience}
}
Copyright German Cancer Research Center (DKFZ) and contributors.
Please make sure that your usage of this code is in compliance with its
license.
This work was partially funded by the Research Campus M2OLIE, which was funded by the German Federal Ministry of Education and Research (BMBF) within the Framework “Forschungscampus: Public-private partnership for Innovations” under the funding code 13GW0388A.