sea-comet/Graduation

Graduation Project: 3D MRI Brain tumor segmentation
(Dong Wanqi)

Report & PPT

Report and PPT and be found here:

• Report
• PPT

Environment

• Platform with GPU.
• Code can be run on Google Colab

Fast running on Google Colab

I also prepared colab notebooks that allow you to run the algorithm on Google Colab. Upload the file to Google Colab and then start training & validation or testing

• runnning on Google Colab: Graduation_Google Colab notebook

Installation

pip install -r requirements.txt

Training & Validation

First change your data folder by modifying the dataset path in src/config.py

• BRATS_TRAIN_FOLDERS is for 5-fold training and validation, which is the path to dataset with ground truth label "seg.nii" file.
• BRATS_VAL_FOLDER is for independent validation or testing, which is the path to dataset with no ground truth labels.
  Both Datasets come from BraTS2020.

# Used for 5 fold-Training and Validation
BRATS_TRAIN_FOLDERS = "your-Path_to/BraTS2020_TrainingData/MICCAI_BraTS_2020_Data_Training"
# Used for independent validation or Testing
BRATS_VAL_FOLDER = "your-Path_to/BraTS2020_ValidationData/MICCAI_BraTS_2020_Data_Valdation"

Then, start training:

Run the following instruction in command line tools.

python -m src.train --width 48 --arch Atten_Unet --epochs 150 # Use Atten_Unet

There are other parameters that can be used. More details on the available options for train.py:

python -m src.train -h

Note that the batch size should be set to 1, for data augmentation is done volume by volume.

After training, you will have a runs folder created containing a directory for each run you have done.

• For each run, a yaml file with the option used for the runs will be generated

• A segs folder containing all the generated .nii.gz files for validation samples. These are the segmentation results.

• First drag the orginal nii.gz file from any of the four modalities t1, t1ce, t2, flair into software ITK-Snap[Download link](A specilaized software used for medical data visualization) to show the brain background, then drag the segmentation result .nii file in folder segs into the software ITK-Snap, too, then the segmentation result on the brain can be visualized.

• You can also use the "display nii data.ipynb" file to see the segmentation

The contents in the runs folder and their corresponding explanations.
Many files are created during training and validation containing the records and results.

- src
    - runs
        - 20220407_154427__fold0_etc 
            20220407_154427__fold0_etc.ymal # Used for testing
            - segs
              - BraTS20_Training_ID.nii
              - ....
            model.txt     # printed model structure
            model_best.pth.tar     # best model weights, used for validation and testing
            patients_indiv_perf.csv    # a log of every patient's segmentation performance in training 
            results.csv     # validation results for each patient in each metric
            val.txt    # validation result at each epoch
            events.out.. # Tensorboard log file, containing lots of graph during training 

Testing

• The yaml file is required to perform inference on testing dataset. model_best.pth.tar should also be under the same directory.
• The script to perform testing is test.py. You should use "--config" option to change the path to the yaml file you created in the training procedure yourself!!

Run the test.py

python -m src.test --config "/content/Graduation/runs/×××××.yaml" --devices 0 --mode val 

For other available options:

python -m src.test -h 

Results for Training & Validation & Testing & Segmentation

毕设训练结果.zip[ Can be found in the link below ]

• 链接: 百度网盘
• 提取码: dg8v

如果上面链接失效，请用下面新的链接

• 链接2: 百度网盘2
• 提取码: ls1w

It contains three seperate results for training & validation and testing.

Please mainly look at the results in folder 3 run_Atten Unet_drop 0.2_noise (0.9,1.1)_150 epoch, 3 pred_Atten Unet_drop 0.2_noise (0.9,1.1)_150 epoch, and 3 Tensorboard 图像 because they use the ultimate chosen parameters and model. The results in folder with a name start with 4 used U-Net structure as comparison, while the results in folder with a name start with 5 used no channel dropiing.

The explanation of each file in folder 3 run_Atten Unet_drop 0.2_noise (0.9,1.1)_150 epoch can be found in Training section, since they are actually the runs folder

Model

Model proposed in this project: Atten_Unet.
It is a varient from 3D-Unet + CBAM block. Each block was modified.
The architecture of Atten_Unet:

Results

Segmentation results:
Here is the comparison between the predicted segmentation result and the ground truth labels for patient No. 9

predicted segmentation result

ground truth labels

They look very similar, aren't they?
The qualitative results are also excellent. Please see the detailed explanation in the report.