This is a practical implementation of a deep learning-based pipeline for automated pre- and postoperative glioma segmentation.

Accurate segmentation of tumor sub-regions (enhancement, edema) can offer the basis for quantitative image analysis towards precision medicine and improvements in individual prognostication. Bringing such tools to clinical reality requires thoughtful implementation. We describe the components required towards clinical deployment of a deep learning-based glioma segmentation model.

We create a pipeline which can route brain tumor DICOM, automate the segmentation algorithm, and push quantitative results back for clinical viewing. This pipeline is composed of 2 servers:

• Data Transit Server: Receive brain tumor DICOM + filter the right input series + push processed DICOM back to PACS
• Inference Server: This server is composed of 3 modules, i.e. (1) pre-processor (2) segmentation (3) post-processor. The deep learning model used in the pipeline can be found here:https://github.com/abenpy/ARCNet. A result incorporating segmentation visualization and volumetrics is routed to the clinical PACS environment.

• Time need for one study: The total time needed between the original scanned DICOM and the processed DICOM presented in PACS is about ~10-15 minutes, depends on the scanner/MRI machine/study itself. Data Routing: 1-2 mins, Preprocessing: 4-6 mins, Segmentation:1-2 mins, Postprocessing:1-2 mins.

We deployed 4 docker containers with daily logs:

• Pre-OP Receiver: receive DICOM using “pynetdicom”

  • Rules: this docker will receive any DICOM sending to it. So the rules should be set from the sending side, i.e the Vendor Neural Archive, which is not included in this git repository. The DICOM series should be like BraTs inputs(https://www.med.upenn.edu/sbia/brats2018/data.html). We suggested to changing is accordingly due to the different naming convention in different organizations. Rules for pre-op receiver in our organisation is: brain tumor protocol; study descriptions ends in Head; study descriptions contains “AX T1 PRE”, “AX T2”, ”SAG 3D FLAIR”, ”SAG MPR” etc.
  • Due to the storage limitation, currently keep only one week original DICOM files.
  • Usually it takes less than 30s to receive a series, e.g. 192 DICOMs for mprage("t1ce" in BraTs convention) needs about 30s to finish routing. However, due to the scanner or other unkown reason, it may take more than 30s, most extremly cases is about 30 mins.

• Post-OP Receiver: receive DICOM using “pynetdicom”

  • Rules: same logic as pre-op receiver. Rules for pre-op receiver in our organisation is:brain tumor protocol; study descriptions ends in Head; study descriptions contains “PERFUSION”(a key series to identify the follow-up/post-op studies), “AX T1 PRE”, “AX T2”, ”SAG 3D FLAIR”, ”SAG MPR” etc.
  • Due to the storage limitation, currently keep only one week original DICOM files.
  • Usually it takes less than 30s to receive a series, e.g. 192 DICOMs for mprage("t1ce" in BraTs convention) needs about 30s to finish routing. However, due to the scanner or other unkown reason, it may take more than 30s, most extremly cases is about 30 mins.

• DICOM Filter:

  • Setting filter rules to select studies with all the 4 input series only, and rename it in BraTs convention, i.e. t1ce, t1, t2, flair. Exclude study which misses any of t1ce, t1, t2, flair.
  • Because it is unknown which series of a study will be sent first, and when all the series will be completely routed, this filter will loop over the DICOM received by pre-op and post-op receiver every 2 minutes.

• DICOM Sender: push processed DICOM back to PACS, using “dcm4che”.(currently was encapusated in Inference Server)

This is the view of segmentation using "t1ce" and "flair" as background, with volumetric report regarding to different part of tumor.

• Docker
• GPU for Inference Server

• Pre-OP Receiver:
  • docker-compose build brats_receive_mri/docker/preop/docker-compose.yml
  • docker-compose up -d
• Post-OP Receiver:
  • docker-compose build brats_receive_mri/docker/postop/docker-compose.yml
  • docker-compose up -d
• DICOM Filter:
  • docker-compose build brats_filter/docker/docker-compose.yml
  • docker-compose up -d
• Inference Server:
  • generate Dockerfile: bash brats/docker/generate_dockerfile.sh. Using neurodocker:0.7.0 to generate a customized Dockerfile which including "fsl", "ants", "dcm2niix". Refer to https://github.com/ReproNim/neurodocker to get more information about "neurodocker".
  • docker-compose build brats/docker-compose.yml
  • docker-compose up -d

• 0.1
  • Initial Release

If you have any questions regarding the code, please contact dongshangshi@gmail.com or raise an issue on the github repo.

Contributors names and contact info

Ben Zhang [dongshangshi@gmail.com]

Inspiration, code snippets, etc.

• CACNN
• Third Party implementation of 1st BraTs 2018
• neurodocker

Copyright (c) 2021, NYU Grossman School of Medicine. All rights reserved.