Deploying Deep Learning-Based Image Segmentation Models Via CERR
Presentations
TU-IePD-TRACK 4-6 (Tuesday, 7/27/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: Despite the availability of several general-purpose platforms for deep learning-based segmentation (DLS), we lack similarly comprehensive platforms with functionality specific to medical images. We present an open-source extension to the Computational Environment for Radiological Research (CERR) for deploying DLS models that leverages CERR’s existing functionality for radiological data import, transformation, management, and visualization.
Methods: CERR’s deep learning pipeline utilizes Java Script Object Notation (JSON) format configuration files to specify input and output data formats; methods and parameters for transforming images, populating channels, and post-processing segmentations. Several pre-processing operations specific to the medical imaging domain are supported, including automated cropping around patient outline, 2D/3D resampling, 2D/2.5D/3D and multimodal channels, and orientation transformations (transverse, sagittal, coronal). CPU and GPU implementations of pre-trained DLS models are packaged using Singularity containers (Linux) and Conda environment archives (Windows, macOS) for distribution. The returned segmentation labels are imported to CERR for further processing and export to DICOM format. Various post-processing operations are supported to minimize manual editing, including filtering to retain connected component(s) by size or within a user-defined region of interest. Custom pre/post-processing functions and dynamic label-to-structure mapping (where the number output labels is variable) are supported through the use of flexible JSON-format configuration files. This pipeline has additionally been made compatible with GNU Octave for license-free use.
Results: CERR's DLS pipeline, along with a JupyterLab notebook demonstrating its usage, are distributed as open-source GPL-copyrighted software at https://www.github.com/cerr/CERR. It has been successfully used in conjunction with MIM workflows to deploy DLS models developed in-house for segmenting organs-at-risk on head & neck CT images and T2-weighted MR prostate images in the clinic.
Conclusion: We developed a comprehensive, open-source pipeline for deploying deep learning-based medical image segmentation models. In addition to its clinical applications, this pipeline facilitates reproducible and consistent segmentation across institutions for research purposes.
Funding Support, Disclosures, and Conflict of Interest: Research was supported by the following grants: (1) P30 CA008748/CA/NCI NIH HHS/United States (2) R01 CA198121/CA/NCI NIH HHS/United States
ePosters
Keywords
Software, Segmentation, Image Processing
Taxonomy
IM/TH- image Segmentation: General (Most aspects)
Contact Email