DeepMedic+

Deep learning for brain metastasis detection and segmentation in longitudinal MRI data

Keywords: Brain Metastasis Segmentation; Brain Metastases Segmentation; Brain Tumor Segmentation; Deep Learning; Machine Learning; Ensemble; Loss Function; Sensitivity Specificity

This is the official code repository for our Medical Physics paper "Deep learning for brain metastasis detection and segmentation in longitudinal MRI data": https://doi.org/10.1002/mp.15863

The backbone of our implementation is from the original DeepMedic repository. Therefore, almost all the instructions from the original DeepMedic repository still apply to our implementation.

Graphical abstract

Citation

To use these codes, please cite our paper:

[1] Y. Huang, C. Bert, P. Sommer, B. Frey, U. Gaipl, L. V. Distel, T. Weissmann, M. Uder, M. A. Schmidt, A. Dörfler, A. Maier, R. Fietkau, and F. Putz, “Deep learning for brain metastasis detection and segmentation in longitudinal mri data,” Medical Physics, vol. 49, no. 9, pp. 5773–5786, 2022.

Latex Bibtex:

 @article{huang2022deep,  
 author = {Huang, Yixing and Bert, Christoph and Sommer, Philipp and Frey, Benjamin and Gaipl, Udo and Distel, Luitpold V. and Weissmann, Thomas and Uder, Michael and Schmidt, Manuel A. and Dörfler, Arnd and Maier, Andreas and Fietkau, Rainer and Putz, Florian},  
 title = {Deep learning for brain metastasis detection and segmentation in longitudinal MRI data},  
journal = {Medical Physics},  
volume = {49},  
number = {9},  
pages = {5773-5786},  
keywords = {brain metastasis, deep learning, ensemble, loss function, MRI, sensitivity specificity},  
doi = {https://doi.org/10.1002/mp.15863},  
url = {https://aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.15863},  
eprint = {https://aapm.onlinelibrary.wiley.com/doi/pdf/10.1002/mp.15863},  
year = {2022}  
}

This repository contains our DeepMedic+ implementation with prior volumes and proposed volume-level sensitivity-specificity (VSS) loss, as well as the pretrained high-sensitivity and high-precision model weights. For retraining on your own datasets, our preprocessing pipeline is also shared.

VSS loss

The VSS loss is located at here.

Pretrained models

Pretrained models are available upon request for multicenter collaboration. Put the pretrained models at the following folder: DeepMedicDualPath\examples\output\saved_models\pretrainedModels

The high sensitivity model is: deepMedicWide1.high_sensitivity.model.ckpt
The high precision model is: deepMedicWide1.high_precision.model.ckpt

Data structure for inference/test

Put the main test volume paths in .\examples\configFiles\deepMedicPlus\test\testChannels_t1c.cfg

put the corresponding prior volume paths in .\examples\configFiles\deepMedicPlus\test\testPriorChannels_t1c.cfg

If no prior volume, simply put the path to a volume with Zero values only. The zero-value volume we used is shared here named "!!AAZero_Volume.nii.gz".

Data preprocessing pipeline

In order to directly apply our pretrained models, your test data should have the same preprocessing pipeline as ours. Hopefully your MRI sequence is also T1 MPRAGE or other similar T1 sequences. Otherwise, the performance of using our pretrained models will be suboptimal, due to different image styles.

Command line for inference/test with pretrained models

python deepMedicRun -model ./examples/configFiles/deepMedicPlus/model/modelConfig_wide1_deeper.cfg -test ./examples/configFiles/deepMedicPlus/test/testConfig.cfg -load ./examples/output/saved_models/pretrainedModels/deepMedicWide1.high_sensitivity.model.ckpt -dev cuda0

python deepMedicRun -model ./examples/configFiles/deepMedicPlus/model/modelConfig_wide1_deeper.cfg -test ./examples/configFiles/deepMedicPlus/test/testConfig.cfg -load ./examples/output/saved_models/pretrainedModels/deepMedicWide1.high_precision.model.ckpt -dev cuda0

For Windows users, you can also use runCMDs.py to run multiple commands.

Training new models for your own data

Preprocess the data with either your own preprocessing pipeline or ours.
Put your training data paths in the config files of .\examples\configFiles\deepMedicPlus\train. Put the volumes, labels, brain masks, and prior volumes into trainChannels_t1c.cfg, trainGtLabels.cfg, trainRoiMasks.cfg, and trainPriorChannels_t1c.cfg, respectively.
In trainConfigwide.cfg, Line 62: please add a new numberOfEpochs value.
Change the $\alpha$ parameter mannualy in cost_functions.py, for example, 0.995 for high sensitivity and 0.5 for high precision. (Sorry that haven't made it available in the args input yet).

Run the command for training

python deepMedicRun -model ./examples/configFiles/deepMedicPlus/model/modelConfig_wide1_deeper.cfg -train ./examples/configFiles/deepMedicPlus/train/trainConfigwide.cfg  -dev cuda0

For fast training, you can fine tune our pretrained model with your own training data with the following command:

python deepMedicRun -model ./examples/configFiles/deepMedicPlus/model/modelConfig_wide1_deeper.cfg -train ./examples/configFiles/deepMedicPlus/train/trainConfigwide.cfg  -load ./examples/output/saved_models/pretrainedModels/deepMedicWide1.high_sensitivity.model.ckpt -dev cuda0

python deepMedicRun -model ./examples/configFiles/deepMedicPlus/model/modelConfig_wide1_deeper.cfg -train ./examples/configFiles/deepMedicPlus/train/trainConfigwide.cfg  -load ./examples/output/saved_models/pretrainedModels/deepMedicWide1.high_precision.model.ckpt -dev cuda0

In trainConfigwide.cfg, Line 62: please add a new numberOfEpochs value, for example, 80, then additional 30 epochs will be trained.

Our System Environment

Tensorflow-GPU version:2.4.1, numpy version:1.19.5 pandas version 1.3.4 it seems that there is compatibility problems with the latest numpy versions.

We run our codes in a Windows system with Python 3.7, CUDA 11.3; Higher versions should be okay.

Acknowledgments

Our implementation is mainly based on DeepMedic.

YixingHuang / DeepMedicPlus

readme