Author: Aditya P. Apte, Joseph O. Deasy, Sharif F. Elguindi, Aditi Iyer, Jue Jiang, Eve Marie LoCastro, Jung Hun Oh, Amita Shukla-Dave, Harini Veeraraghavan 👨🔬
Affiliation: Department of Medical Physics, Memorial Sloan Kettering Cancer Center, Memorial Sloan Kettering Cancer Center 🌍
Purpose: We present publicly shareable applications (apps) for AI-based radiotherapy segmentation workflows with pyCERR on Seven Bridges Cancer Genomics Cloud-based platform (CGC-SB)
Methods: Running AI-based models typically requires customized software configurations and GPU hardware. Seven Bridges Cancer Genomics Cloud by Velsera (CGC-SB) is a NCI-funded online platform for cancer research. CGC-SB provides data storage, interactive computational environments, ability for sharing analysis workflows and access to powerful computational instances.
Users create and share analysis workflows as “apps” launched in custom-configured Docker environments. We created apps, defined in Common Workflow Language (CWL) for AI segmentation of normal tissue organs-at-risk (OAR’s). These AI models have been clinically implemented in our institution over the last 5 years. Segmentations are available for lung, cardiac subtruces, and head-and-neck.
App segmentation jobs can be initiated in CGC-SB as tasks in the cloud from within CGC-SB web portal. For optimal inclusion in research imaging workflows these tasks can also be initiated from users’ local clients (workstations) via REST calls, or API wrappers for Python or R.
Our notebooks employ pyCERR, is the newly-developed port of CERR radiotherapy analysis platform, including extensions for image conversion, auto-segmentation, and radiomics analysis.
Results: We have implemented the pyCERR auto-segmentation routines using GPU as apps on CGC-SB, providing accurate rapid segmentation for organs-at-risk including heart, lung, head and neck structures. We provide Jupyter notebook round-trip to segment images on CGC, demonstrating how to initialize, retrieve and review results from their local client. We aim to expand our CGC-SB framework to allow users to rapidly perform large-scale network training.
Conclusion: pyCERR AI segmentation routines are available to CGC-SB users to quickly segment radiotherapy data without requiring local setup beyond basic command line tools to remotely launch CGC apps.