Skip to content

AICONSlab's DL benchmarking platform to OOD data in MRI

License

Notifications You must be signed in to change notification settings

AICONSlab/roodmri

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ROOD-MRI: Benchmarking the Robustness of deep learning segmentation models to Out-Of-Distribution data in MRI

image

Getting started

The examples folder contains example scripts for working with the DatasetGenerator class and calculate_metrics function. We will be adding more examples to this folder, stay tuned!

The basic workflow is:

  1. Generate a benchmarking dataset from a pre-existing test set using the DatasetGenerator class.
  2. Evaluate a trained model on the benchmarking dataset, generating a dataframe or .csv file containing segmentation metrics for each sample.
  3. Calculate benchmarking/robustness metrics from the dataframe/.csv generated in step 2 using calculate_metrics.

Dependencies for this library are listed in the requirements.txt file (including MONAI and TorchIO).

1. Generate a benchmarking dataset

Skip this step if you're using a pre-existing benchmarking dataset (see links to existing datasets below).

If you have a dataset directory that looks like this:

/home/user/data/
|-- train_data
`-- test_data
    |-- case_01
    |   |-- t1.nii.gz
    |   `-- seg_label.nii.gz
    |-- case_02
    |   |-- t1.nii.gz
    |   `-- seg_label.nii.gz
    .
    .
    .
    `-- case_99
        |-- t1.nii.gz
        `-- seg_label.nii.gz

First, glob the files into a structured list:

from pathlib import Path

data_dir = Path('/home/user/data/')
image_paths = [str(path) for path in sorted(data_dir.glob('test_data/*/t1.nii.gz'))]
label_paths = [str(path) for path in sorted(data_dir.glob('test_data/*/seg_label.nii.gz'))]
input_files = [{'image': img, 'label': lbl} for img, lbl in zip(image_paths, label_paths)]

Then, run the DatasetGenerator over the input files:

from roodmri.data import DatasetGenerator

out_path = '/home/user/data/benchmarking'   # specify the path to put benchmarking samples

generator = DatasetGenerator(input_files, out_path)
generator.generate_dataset()
generator.save_filename_mappings(Path(out_path) / 'filename_mappings.csv')   # save new filename mappings

The folder specified by out_path will now be populated with sub-folders named Affine_1, Affine_2, ..., RicianNoise_4, RicianNoise_5, ... containing transformed samples from the test set. In the name RicianNoise_4, RicianNoise refers to the transform applied and 4 refers to the severity level. The image below illustrates an example of the five default severity levels on a sample T1-weighted image for (a) ghosting, (b) isotropic downsampling, and (c) MRI (Rician) noise:

image

For more details and examples using different initial directory structures, see the examples/dataset folder.

2. Evaluate your model(s) on the benchmarking dataset

The end result of this step should be a csv file or dataframe with segmentation results for each benchmarking sample, as well as the original clean test set:

Model,Task,Transform,Severity,Subject_ID,DSC,HD95
unet_a,WMHs,Affine,1,Subject_001,0.82,1.41
unet_a,WMHs,Affine,1,Subject_002,0.79,2.34
.
.
unet_a,WMHs,Clean,0,Subject_001,0.85,1.41
.
.
unet_f,Ventricles,Affine,1,Subject_001,0.90,1.56
.
.

Since users' own evaluation pipelines may vary significantly (pre-processing, transforms, dataloaders, etc.), we do not provide modules to evaluate models on the benchmarking dataset. Rather, we suggest that users use their own existing pipelines to generate a csv file such as the one above. We will be uploading some of our own examples to the examples folder, including code for how to parse the transform/severity level folder name.

For more details regarding the requirements for the csv/dataframe, see metric_calculations.py in the examples folder.

3. Calculate benchmarking metrics

After producing a csv/dataframe with segmentation results, you can use the calculate_metrics function to generate a suite of benchmarking metrics:

from pathlib import Path

import pandas as pd

from roodmri.metrics import calculate_metrics

data_path = '/home/user/data/model_evaluation_results.csv'   # change to location of csv
save_path = '/home/user/benchmarking/'   # change to desired location of output files
df = pd.read_csv(data_path)
transform_level_metrics, aggregated_metrics = calculate_metrics(
    df=df,
    transform_col='Transform',
    severity_col='Severity',
    metric_cols={'DSC': True, 'HD95': False},
    clean_label='Clean',
    grouping_cols=['Model', 'Task']
)
transform_level_metrics.to_csv(Path(save_path) / 'transform_level_metrics.csv')
aggregated_metrics.to_csv(Path(save_path) / 'aggregated_metrics.csv')

The image below demonstrates an example of using benchmarking metrics to comparing model architectures. The numbers in the lower- and upper-left corners of the top-row and bottom-row subplots, respectively, correspond to the mean degradation for each model (top row: Dice similarity coefficient; bottom row: modified (95th percentile) Hausdorff distance):

image

For more documentation, see metric_calculations.py in the examples folder, or calculate.py which contains the calculate_metrics function. For metric formulations and how to use them, check out our paper.

Links to existing datasets

See the list below for download links to existing benchmarking datasets:

Credits

If you like this repository, please click on Star!

If you use this package for your research, please cite our paper:

Boone, L., Biparva, M., Forooshani, P. M., Ramirez, J., Masellis, M., Bartha, R., ... & Goubran, M. (2022). ROOD-MRI: Benchmarking the robustness of deep learning segmentation models to out-of-distribution and corrupted data in MRI. arXiv preprint arXiv:2203.06060.

@article{boone2022rood,
  title={ROOD-MRI: Benchmarking the robustness of deep learning segmentation models to out-of-distribution and corrupted data in MRI},
  author={Boone, Lyndon and Biparva, Mahdi and Forooshani, Parisa Mojiri and Ramirez, Joel and Masellis, Mario and Bartha, Robert and Symons, Sean and Strother, Stephen and Black, Sandra E and Heyn, Chris and others},
  journal={arXiv preprint arXiv:2203.06060},
  year={2022}
}

This work has been enabled by the flexibility and modularity of the MONAI and TorchIO libraries. If you like our work and aren't familiar with theirs already, go check them out!