Grand Challenges ISLES 2022 Evaluation Container

This repo contains the evaluation container for the ISLES 2022 challenge (e.g. ATLAS). The build.sh, test.sh, export.sh scripts along with msot of Dockerfile were generated by evalutils version 0.3.1. The code presented here is not needed by participants for the challenge, but it is made available for transparency and to give participants a way to test the evaluations on their own machines.

The Dockerfile expects ground truth labels to be in the ground-truth directory. The format must be a BIDS dataset that conforms to the user-supplied settings in settings.py.

The settings file (settings.py) can be used to control data loading, which scoring functions are used, and which summary statistics are returned.

Adapting this code to other BIDS datasets used in other challenges should require only the following settings:

• Updating GroundTruthBIDSDerivativeName and PredictionBIDSDerivativeName to your derivative names.
• Updating GroundTruthEntities and PredictionEntities to values matching your data. See the BIDSIO documentation for more information on how to use these.
• Importing your desired metrics into settings.py and adding them to the ScoringFunctions dictionary.

Setting Tutorial

Data Setting

Provided that the requested Test dataset is located in the correct folder (external_dataset/TaskXXX_MYTASK).
The modalities (ADC, DWI) in the test dataset used must be named in the correct format:
external_dataset/TaskXXX_MYTASK/imagesTs/XXX_0000_0000.nii.gz: for ADC dataset
external_dataset/TaskXXX_MYTASK/imagesTs/XXX_0000_0001.nii.gz: for DWI dataset
external_dataset/TaskXXX_MYTASK/labelsTs/XXX_0000.nii.gz: for Label dataset (if exist)

Inference

The inference step consists of running the following sequence of code:
bash run.sh TaskXXX_MYTASK 4: code for DWI with transfer learning & multi-tasking
bash run.sh TaskXXX_MYTASK 6: code for ADC+DWI with transfer learning
bash run.sh TaskXXX_MYTASK 10: code for ensemble

Weights trained by our proposed method

• Download: OneDrive
• 3d_fullres_dwi_tf_mt.zip: Single-modality(DWI) model weight with trainsfer & multi-task learning
• 3d_fullres_adc_dwi_tf_t2.zip: Multi-modality(ADC+DWI) model weight with transfer learning
• The trained models should be located in the nnUNet_model/ folder