DeepResection

Deep Learning-Based Automated Segmentation of Resection Cavities on Postsurgical Epilepsy MRI

T. Campbell Arnold*, Ramya Muthukrishnan*, Akash R. Pattnaik, Adam Gibson, Nishant Sinha, Sandhitsu R. Das, Brian Litt, Dario J, Englot, Victoria L. Morgan, Kathryn A. Davis, Joel M. Stein

*These authors contributed equally

Center for Neuroengineering and Therapeutics, University of Pennsylvania

Deep learning code for neurosurgery resection zone segmentation on T1 MRI, implemented in Keras with Tensorflow backend.

Prerequisites

Linux OS / Ubuntu
Python 3.6
Python package dependencies in requirements.txt
Postoperative (and preoperative) MRI in NIfTI format

Getting Started

Clone the repo: git clone https://github.com/penn-cnt/DeepResection.git.

Enter the DeepResection directory. Make a model directory within the repo using the command mkdir model. Then, download the 3 model weights at this link and save them in the model directory as model/model_axial.h5, model/model_coronal.h5, and model/model_sagittal.h5.

Running the Pipeline

First, please ensure that all input images are in LAS orientation, otherwise resection predictions are likely to be inaccurate.

Run segmentation only:

./pipeline/resection_segmentation_only_pipeline.sh patient_name postop_mri.nii output_dir

Run entire pipeline, including volumetric resection report:

./pipeline/resection_pipeline.sh patient_name preop_mri.nii postop_mri.nii output_dir

Run pipeline with deformable registration:

./pipeline/resection_deformable_pipeline.sh patient_name preop_mri.nii postop_mri.nii output_dir

Example outputs

Predicted mask

After running the pipeline, the predicted masks in each dimension should be named predicted_mask_axial.nii.gz, predicted_mask_coronal.nii.gz, and predicted_mask_sagittal.nii.gz in the specified output directory. The final predicted mask using majority voting (with 2 votes) is named predicted_mask.nii.gz in the output directory. All 4 masks can be opened in an image viewer, such as ITK-Snap, alongside the postoperative input.

Volumetric resection report

If running the full pipeline, the numeric volumetric resection results should be in output_dir/resected_results.txt, the HTML report should be in output_dir/resection_report.html, and a visualization of the resection should be in output_dir/resection_views.png. Examples are shown below.

Text file output:

Total resection volume (cubic cm): 19.32525634765625
Frontal_Inf_Orb_R: 99.947% remaining
Hippocampus_R: 91.254% remaining
ParaHippocampal_R: 45.888% remaining
Amygdala_R: 81.513% remaining
Fusiform_R: 79.39699999999999% remaining
Temporal_Sup_R: 99.963% remaining
Temporal_Pole_Sup_R: 68.662% remaining
Temporal_Mid_R: 92.673% remaining
Temporal_Pole_Mid_R: 41.357% remaining
Temporal_Inf_R: 82.94% remaining
Cerebelum_3_R: 99.81% remaining

HTML report:

Resection views PNG output:

Web App

A web application is also provided for easy and interactive clinical use. The home page consists of two pipelines: Resection Mask and Volumetric Report.

The resection mask page allows you to select if the mask is continuous and upload the post-operative MRI. The patient ID is used for labeling files and a drop down menu gives the option of different atlases for the report.

If the DeepResection model has already been applied and the mask is given, then the Volumetric Report pipeline generates the report.

The report consists of a table with regions from the selected atlas and estimated resected volume. The widget allows for the user to scan through cross sections of the post-operative MRI and the mask is superimposed in blue. An option to save a PDF of the report is given.

An option for the user to give feedback is given. This feedback will be used to improve the DeepResection pipeline.

tcama/DeepResection