Authors

• Development: Milad Sadeghi DM

Sample Pipeline

Pre-processing:

Desired dataset structure:
Input data shape: 64x64x10

.
└── Train
    ├── 100
    │   ├── mask_100.nii
    │   ├── MD_100.nii
    │   ├── rCBV_100.nii
    ├── 101
    │   ├── mask_101.nii
    │   ├── MD_101.nii
    │   ├── rCBV_101.nii

Do preprocessing on dataset:

python preprocessing.py --verbose --dataset <path to dataset directory>

This will create a new directory called npy_files, then creates specific folders for each sample we have in dataset and saves image.npy and mask.py files in these folders.

• Output data shape:
  • Image: 64x64x16x2
  • Mask: 64x64x16x3

npy_files/
├── 0
│   ├── mask_0.npy
│   └── image_0.npy
├── 1
│   ├── mask_1.npy
│   └── image_1.npy

Training the model:

python train.py --verbose --dataset <path to npy_files> --learning_rate 0.0001 --batch_size 2 --epochs 100 --save ./

After completing training process, command above will save entire trained model as a .pb file , save training process history details as a .csv file and plot training process diagrams.

Build report file:

As we mentioned earlier, train script saves training process history details. The reporter.py file drops unnecessary rows and columns and calculates f1_score value for both training and validation.

python reporter.py -v --file_path <path to log_file.csv> --save <path to save new .csv file>

Results

Train:

Fold	Accuracy	IOU Score	Loss	Precision	Recall	F1-Score
1	0.968775	0.897440	-1.903993	0.968788	0.968766	0.968777
2	0.970905	0.903380	-1.730110	0.970907	0.970904	0.970905
3	0.972417	0.907598	-1.769276	0.972422	0.972415	0.972419
4	0.972910	0.910098	-1.937584	0.972916	0.972903	0.972909
5	0.973350	0.912430	-1.882656	0.973355	0.973342	0.973349
6	0.972398	0.907724	-1.793769	0.972407	0.972395	0.972401
7	0.973002	0.905900	-1.848724	0.973010	0.972997	0.973004
8	0.971803	0.906942	-1.939411	0.971807	0.971801	0.971804

Validation:

Fold	Accuracy	IOU Score	Loss	Precision	Recall	F1-Score
1	0.912868	0.674202	-1.214981	0.912912	0.912853	0.912882
2	0.908874	0.599501	-0.614096	0.908877	0.908874	0.908876
3	0.900650	0.637642	-0.932565	0.900656	0.900642	0.900649
4	0.903141	0.668193	-1.140507	0.903154	0.903133	0.903143
5	0.890136	0.610992	-0.862575	0.890166	0.890125	0.890145
6	0.893863	0.612036	-0.745553	0.893863	0.893863	0.893863
7	0.895732	0.613254	-0.769792	0.895738	0.895725	0.895732
8	0.885597	0.618338	-0.897051	0.885608	0.885578	0.885593

Making a prediction

First you have to do preprocess on your sample different modalities and generate image.nii and mask.npy and then run predict.py script:

python predict.py --model <path to saved model directory> --image <path to preprocessed .npy image file>
--mask <path to preprocessed .npy mask file> --verbose --save_plot --slice 6

Sample output:

License

This project licensed under the MIT License - see the LICENSE file for more details.

References

1. Optimized U-Net for Brain Tumor Segmentation
2. Optimized High Resolution 3D Dense-U-Net Network for Brain and Spine Segmentation
3. A novel fully automated MRI-based deep-learning method for classification of IDH mutation status in brain gliomas