From Pixels to Patients: Practical Pipelines for Implementing AI in Radiation Oncology
Description
Artificial intelligence holds significant promise for radiation oncology, yet clinical adoption is frequently stalled by the gap between conceptual understanding and practical implementation. This three-part educational track provides a start-to-finish playbook for building robust, reproducible AI pipelines, from raw data through deployment. The first session addresses the foundational challenge of data extraction and curation. Radiation oncology data is locked inside complex DICOM objects spanning Imaging, RTSTRUCT, RTPLAN, and RTDOSE. Attendees will learn reproducible strategies for converting these into AI-ready formats such as NIfTI and NumPy arrays. The second session moves from curated datasets into model development using PyTorch. Practical walkthroughs cover construction of efficient 3D/4D dataloaders, RT-specific augmentation strategies, and reproducibility best practices including dataset splitting, version control, and leakage prevention, common pitfalls that can silently undermine model validity. The third session confronts the often-underappreciated challenge of clinical deployment. Drawing on real-world institutional experience, this session candidly examines deployment failures, edge cases, and gaps in post-deployment monitoring. Attendees will leave with a practical QC framework for input validation, output sanity checking, drift detection, and ongoing governance applicable to clinics at any stage of AI adoption. Attendees will receive reference code, open notebooks, and templates to accelerate their own AI initiatives. Basic Python familiarity is beneficial but not required for the full track.