12.3. Multi AI Pipeline


Multi AI pipeline takes a single volume and splits them into multiple ROIs, these ROIs are then fed into respective AI operators. Results from AI operators are merged into a single volume.

Pipeline uses roi-generator operator for splitting data into rois and merging data back into a single volume. Liver tumor, Lung tumor, Colon tumor and Spleen AI operators are used in the ROI configuration.

Multi AI pipeline has 4 AI models, it needs 24 GB of GPU memory with the example test dataset.

Data required for running this pipeline is stored in artifactory.

  • Get the data packaged with pipeline from NGC.

  • Unzip app_multiai-input_v1.zip file. There are 4 folders inside multiAIData folder namely ‘data’, ‘roiZ’, ‘roiXYZ’ and ‘render’

  • If roi in Z dimensions is desired then copy roiZ/roi.txt to data/roi.txt, if ROI in XYZ dimensions is desired then copy roiXYZ/roi.txt to data/roi.txt.

  • If results is desired in render server, copy render/config_render.json to data/config_render.json

  • Look at /clara/common/models folder. Spleen, LungTumor, ColonTumor, LiverTumor models must be present.

  • If any of the model is missing, get the models from NGC as zip folder.

  • Unzip app_multiai-model_v1.zip file. Copy missing models from zip file to /clara/common/models

The following are the execution steps for multi AI Pipeline:

  • Data preparation

    • Store location of folder ‘data’ after updating configuration files as described in above section.

  • Platform Installation: Ensure that platform is installed successfully and Clara CLI is installed correctly.

  • Required Models: Ensure that spleen, colon tumor, lung tumor and liver tumor models are present in /clara/common/models folder.

  • Required Operators: Following operators must be installed. If these operators are not present in the system, either build them locally or pull them from https://ngc.nvidia.com/containers.

    • roi-generator

    • app_lungtumor

    • app_livertumor

    • app_spleen

    • app_colontumor

    • register-results

    • pod-manager

    Update the pipeline definition with appropriate container image and tag information. Ensure that container image and tag are present in the system before executing the pipeline.

  • Create roi-generator: This is not a mandatory step. Alternatively, ‘roi-generator’ can be build locally. Following are the steps:

    • Execute ‘docker images’ and locate the ‘TAG’ of clara/python-base image.

    • Execute following to tag python-base image, use value of ‘TAG’ as identified in above step:


      docker tag clara/python-base:'TAG' clara/python-base:latest

    • Navigate to Applications/Operators/image_processing/app_roi/ folder in sdk. Execute ‘make’.

    • Execute ‘docker images’ and verify that roi-generator operator is created. Update the pipeline definition with appropriate container image ang tag information

  • Execute following in sequence using Clara cli:


    clara create pipeline -p <path to multiAI pipeline> clara create jobs -n <name> -p <pipeline ID> -f <path to folder 'data' as identified in step 1> clara start job -j <JOB ID>

  • Check payloads for the mask or visualize the results in render server.

© Copyright 2018-2020, NVIDIA Corporation. All rights reserved.. Last updated on Feb 1, 2023.