Sequence-Agnostic Model with Cross-Sequence Distillation for Localization of Seizure Onset Zone

This is the official PyTorch implementation of the paper Non-invasive Localization of Seizure Onset Zone using Clinically Acquired MRI in Children with Drug-Resistant Epilepsy: a Sequence-Agnostic Model with Cross-Sequence Distillation that was submitted to Medical Image Analysis journal and is currently under review.

Requirements

• Python >= 3.9
• PyTorch >= 1.12.0
• torchmanager >= 1.1.0
• Monai >= 1.1
• Multimodality
• Download the script from the above Multimodality branch in your code directory - since it will be used in the main code.

Get Started

The following steps are required to replicate our work:

1. Dataset

   • This project uses a private clinical dataset of MRI scans from children with drug-resistant epilepsy.
   • The dataset basically consists of features derived from various MRI sequences such as T1w, T2w, FLAIR, DWI and DWIC, for every node of the brain of children with Drug-Resistant Epilepsy.
   • For access to similar datasets for research purposes, please contact the authors.

2. Data preprocessing

   • The SMOTE augmentation is performed by augment_single_node_left_hemis_part2.py and augment_single_node_right_hemis_part2.py for the nodes of the left and right hemispheres respectively.
   • The data loading steps are provided in the data/dataset_ez.py file. This file uses the SMOTE augmented data.
   • It performs all the necessary preprocessing steps to prepare the MRI data for training and validation.

Training

1. Refer to the training configuration in train_right.py for the default hyper-parameters to train the models for the right hemisphere nodes:

# To train the model on nodes of the right hemisphere
./train_ALL_right.sh

2. Refer to the training configuration in train_left.py for the default hyper-parameters to train the models for the left hemisphere nodes:

# To train the model on nodes of the left hemisphere
./train_ALL_left.sh

3. Please note that depending on the node numbers, we may need to change the shell script name. This is because the training on all the nodes is sequential and we used multiple shell scripts to obtain faster results.

4. The training process can be customized by modifying the following parameters in the shell scripts:

   • Batch size
   • Learning rate
   • Number of epochs
   • Number of Modality or sequence selection
   • Device selection

Testing

1. The pre-trained models are stored in the experiments/ folder.

   • The pre-trained models are not included in this GitHub repository due to space limitations.

2. Refer to the eval_right.py for the default settings to test the models for the right hemisphere nodes:

# To evaluate all nodes of the right hemisphere
./eval_ALL_right.sh

3. Refer to the eval_left.py for the default settings to test the models for the left hemisphere nodes:

# To evaluate all nodes of the left hemisphere
./eval_ALL_left.sh

4. Please note that depending on the node numbers, we may need to change the shell script name. This is because the evaluation on all the nodes is sequential and we used multiple shell scripts to obtain faster results.

5. The testing process can be customized by modifying the parameters in the shell scripts or directly in the eval_left.py and eval_right.py files.

Final Results

1. After evaluating the results, combine the results for all nodes of the left hemisphere with:

python combine_node_excel_sheet_results_eval_left.py

2. After evaluating the results, combine the results for all nodes of the right hemisphere with:

python combine_node_excel_sheet_results_eval_right.py

The above will provide excel files with the results for all nodes.

Citation