University of Texas Southwestern Medical Center Dallas, TX
M. Chen1, J. Lin2, M. Kazemimoghadam1, Q. Wang1, H. Jiang3, Y. K. Park2, M. H. Lin2, A. R. Godley2, A. Pompos2, X. Gu4, and W. Lu1; 1Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, TX,, Dallas, TX, 2Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX, 3NeuralRad LLC, Madison, WI, 4Stanford University Department of Radiation Oncology, Palo Alto, CA
Purpose/Objective(s): Despite advances in medical image segmentation, their integration into radiotherapy workflow is often impeded by a lack of customizability in meeting dynamic and diverse clinical needs. Similarly, existing single-vendor quality assurance (QA) software for treatment plan and delivery often falls short in meeting the efficiency and comprehensiveness required for QA in a dynamic and versatile clinical environment. To address these gaps, we have developed two innovative synergistic platforms: Segmon and Dosemon. Materials/
Methods: Segmon encompasses a comprehensive suite of segmentation models, including popular deep learning networks like DeepMedic, EnU-NET, and nnU-Net, trained to cover diverse anatomical sites and various treatment and imaging protocols. Users can define pre/post processing and model/ensemble selection all through config and task files. This capability promotes versatility and customization, from simple morphological and Boolean operations on structures to intricate templates that incorporate structure lists and contour preferences tailored to specific requirements and standardization of disease-oriented teams or individual physicians. Dosemon leverages an in-house GPU Monte Carlo (MC) dose engine with full electron/photon transport and electromagnetic field modeling. It utilizes a machine-agnostic auto-commissioner based on measurement data to overcome challenges of machine modeling and patient-specific output modeling for performing treatment plan and delivery QA across diverse clinical LINACs. Both platforms share a common foundation, including an SQL database, config and task file-based programming, a WebServer, a WebAssembly (WASM)-based portal, and dynamic task graph (DTG) scheduling. All services run as daemon processes without requiring user intervention. Moreover, their integration with Treatment Planning Systems and Oncology Information Systems ensures a holistic approach to patient care. Results: Segmon provides contours and labels to users for OAR segmentation within 5 min once images are ready. Its usage has doubled in nine months, extending its application to 13 disease sites for CT and 6 for MRI based segmentation. Dosemon computes MC doses and generates QA reports within 2 min, immediately following plan export or treatment completion. On average, Dosemon drastically reduces human involvement, from 15-30 min to <1 min per plan/treatment, while enhancing the scope of checks. Dosemon has been critical for QA of online adaptive treatment plans, as no measurement solution is viable, and commercial software is too slow. Conclusion: Segmon and Dosemon represent robust platforms for AI-driven auto-segmentation and patient-specific QA in radiotherapy. The config and task file-based programming and DTG architecture are key to the platforms’ adaptability, customizability, and extensibility. The synergy between Segmon and Dosemon epitomizes innovation and efficiency within radiotherapy.
