Artificial Intelligence-Based Radiotherapy Plan Review Education

Skylar Slade Gay, BS (Advisor: Laurence E. Court, PhD)

Radiotherapy is an essential component of modern cancer care, but training in treatment plan review remains constrained by the increasing complexity of planning and by limitations of the clinical learning environment. Although artificial intelligence (AI) has improved many radiotherapy workflows, including imaging, contouring, and automated planning, its use in education has remained limited. Traditional apprenticeship-based training in plan review is restricted by limited case variety, competing clinical priorities, delayed feedback, and the high-stakes nature of patient care, all of which reduce opportunities for deliberate practice. These constraints may contribute to persistent gaps in trainee understanding and skill that extend into clinical practice.

The goal of this dissertation was to develop approaches for simulating radiotherapy plans of varying quality and to translate those methods into an interactive educational platform for training in treatment plan review. First, computational techniques were developed to intentionally generate realistic suboptimal dose distributions from high-quality plans. These methods introduced controllable errors that reduced organ-at-risk sparing, decreased target conformality, or created hotspots within targets, thereby producing a broad range of clinically relevant training examples without requiring large datasets of previously identified poor-quality plans. Across head-and-neck and gynecological cases, these simulated plans demonstrated statistically significant dosimetric degradation in the intended regions and were generally judged by experienced clinicians to be sufficiently realistic for educational use.

Next, a deep learning-based 'virtual dosimetrist' was developed to modify radiotherapy dose distributions in response to natural language directives. This cross-modal architecture combined volumetric patient and dose information with text prompts describing requested dose changes. Trained on clinically approved and intentionally replanned head-and-neck cases, the models were able to rapidly generate dose modifications and realistically increase or decrease dose as prompted. On an unseen test set, nearly all predicted dose distributions (98.8%) were equivalent to treatment planning system-generated dose distributions under a predefined equivalence criterion, supporting the realism and fidelity of the approach.

Finally, these methods were integrated into a web-based radiotherapy 'flight simulator' designed to provide a clinic-like, low-stakes environment for plan review training. The platform allowed users to review simulated suboptimal plans, compare them with high-quality reference plans, and practice improving plan quality through interactive directives. In deployment across local multidisciplinary users and an international cohort of radiation oncology residents, use of the platform was associated with plan review assessment gains exceeding 10 percentage points overall, increased user confidence in recognizing and addressing plan deficiencies, and favorable ratings of usability and educational value.

Together, this work establishes a new approach for AI in radiotherapy education, shifting its role from passive automation or information retrieval toward active simulation-based training. By enabling realistic generation and rapid interactive revision of suboptimal plans, these methods provide scalable opportunities for deliberate practice in plan review that are difficult to achieve in routine clinical environments. This work demonstrates that AI can function not only as a clinical assistant, but also as an educational partner capable of expanding access to high-quality radiotherapy training across institutions and resource settings.

Advisory Committee:

• Laurence Court, PhD, Chair
• Carlos Cardenas, PhD
• Tucker Netherton, PhD
• Brent Parker, PhD
• Chelsea Pinnix, MD
• Sanjay Shete, PhD

Join via Zoom (Please contact Mr. Gay for his Zoom meeting info.)