12.6. Hippocampus Segmentation Pipeline
The Hippocampus Segmentation pipeline is one of the reference pipelines provided with Clara Deploy SDK. A pre-trained model for volumetric (3D) segmentation of the hippocampus head and body from a mono-modal MRI image is used in the pipeline. This pipeline depends on the Clara Deploy DICOM Adapter to receive DICOM images, specifically single-channel 3D MRI images. Once the DICOM instances are stored, the pipeline is started to first convert the DICOM instances to a volume image in MetaIO format and then processes the input volume image with the Hippocampus Segmentation AI model. The AI model generates the labeled segmentations (within the hippocampus) as a mask on slices of the volume of the same size as the input. The hippocampus anterior is labeled as 1, the hippocampus posterior is labeled as 2, and the background is labeled as 0. In its final steps, the pipeline outputs the mask as a new DICOM series using the original study instance UID; also, it outputs the original and segmented volumes to the Clara Deploy Render Server for visualization on the Clara Dashboard.
The Hippocampus Segmentation pipeline is defined in the Clara Deploy pipeline definition language. This pipeline utilizes built-in reference containers to construct the following set of operators:
- The dicom-reader operator converts input DICOM data into volume images in mhd format.
- The hippocampus-segmentation operator performs AI inference against the NVIDIA Triton Inference server to generate hippocampus anterior and posterior segmentation volume images.
- The dicom-writer converts the segmented volume image into DICOM instances with a new series instance UID and the original DICOM study instance UID.
- The register-dicom-output operator registers the DICOM instances with the Clara Deploy DICOM Adapter which in turn stores the instance on external DICOM devices per configuration.
- The register-volume-images-for-rendering operator registers original and segmented volume images with the Clara Deploy Render Server for visualization.
The following is the details of pipeline definition, with comments describing each operator’s functions as well as input and output.
api-version: 0.4.0
name: hippocampus-pipeline
operators:
# dicom reader operator
# Input: '/input' mapped directly to the input of the pipeline, which is populated by the DICOM Adaptor.
# Output:'/output' for saving converted volume image in MHD format to file whose name
# is the same as the DICOM series instance ID.
- name: dicom-reader
description: Converts DICOM instances into MHD, one file per DICOM series.
container:
image: clara/dicom-reader
tag: latest
input:
- path: /input
output:
- path: /output
# hippocampus-segmentation operator
# Input: `/input` containing volume image data, MHD format, with a single volume.
# Output: `/output` containing segmented volume image, MHD format.
# `/publish` containing original and segmented volume images, MHD format,
# along with rendering configuration file.
- name: hippocampus-segmentation
description: Segmentation of hippocampus inferencing using DL trained model.
container:
image: clara/ai-hippocampus
tag: latest
requests:
memory: 8192
input:
- from: dicom-reader
path: /input
output:
- path: /output
name: segmentation
- path: /publish
name: rendering
services:
- name: trtis
# Triton Inference Server, required by this AI application.
container:
image: nvcr.io/nvidia/tritonserver
tag: 20.07-v1-py3
command: ["tritonserver", "--model-repository=$(NVIDIA_CLARA_SERVICE_DATA_PATH)/models"]
# services::connections defines how the TRTIS service is expected to
# be accessed. Clara Platform supports network ("http") and
# volume ("file") connections.
connections:
http:
# The name of the connection is used to populate an environment
# variable inside the operator's container during execution.
# This AI application inside the container needs to read this variable to
# know the IP and port of TRTIS in order to connect to the service.
- name: NVIDIA_CLARA_TRTISURI
port: 8000
# Some services need a specialized or minimal set of hardware. In this case
# NVIDIA Tensor RT Inference Server [TRTIS] requires at least one GPU to function.
# dicom writer operator
# Input1: `/input` containing a volume image file, in MHD format, name matching the DICOM series instance UID.
# Input2: `/dicom` containing original DICOM instances, i.e. dcm file.
# Output: `/output` containing the DICOM instances converted from the volume image, with updated attributes
# based on original DICOM instances.
- name: dicom-writer
description: Converts MHD into DICOM instance with attributes based on the original instances.
container:
image: clara/dicom-writer
tag: latest
input:
- from: hippocampus-segmentation
name: segmentation
path: /input
- path: /dicom
output:
- path: /output
name: dicom
# register-volume-images-for-rendering operator
# Input: Published original and segmented volume images, MHD format, along with rendering configuration file
# from hippocampus-segmentation operator.
# Output: N/A. Input data will be sent to the destination, namely `renderserver` for Render Server DataSet Service.
- name: register-volume-images-for-rendering
description: Register volume images, MHD format, for rendering.
container:
image: clara/register-results
tag: latest
command: ["python", "register.py", "--agent", "renderserver"]
input:
- from: hippocampus-segmentation
name: rendering
path: /input
# register-dicom-output operator
# Input: `/input` containing DICOM instances in the named output, `dicom` from dicom-writer operator.
# Output: N/A. Input data will be sent to the destinations, namely DICOM devices, by the Clara DICOM SCU agent.
- name: register-dicom-output
description: Register converted DICOM instances with Results Service to be sent to external DICOM devices.
container:
image: clara/register-results
tag: latest
command: ["python", "register.py", "--agent", "ClaraSCU", "--data", "[\"MYPACS\"]"]
input:
- from: dicom-writer
name: dicom
path: /input
Please refer to the How to Run a Reference Pipeline section to learn how to register a pipeline, configure the DICOM Adapter, and execute the pipeline.