Intentional creation of suboptimal, realistic dose distributions

Area of Science:

• Medical Physics
• Radiation Oncology Education
• Artificial Intelligence in Medicine

Background:

• Radiation oncology residents often lack confidence in radiotherapy plan quality assessment due to limited practice opportunities in clinical settings.
• Current training environments in routine clinical practice do not offer sufficient time or diverse cases for residents to develop expertise in plan review.
• The need for improved resident education in plan review is critical, as plan quality directly impacts patient outcomes.

Purpose of the Study:

• To develop a method for generating realistic, yet controllable suboptimal, radiotherapy dose distributions for educational purposes.
• To create a low-stakes training environment that provides radiation oncology residents with ample case examples for practicing plan review.
• To enhance the curriculum with diverse and controllable suboptimal plans, addressing limitations of smaller programs and rare cancer types.

Main Methods:

• Generated high-quality dose distributions using a pre-trained deep learning model.
• Directly altered dose distributions to create three types of suboptimal plans: reduced organ-at-risk sparing, decreased target conformity, and target hotspots.
• Assessed the realism of generated suboptimal dose distributions through review by experienced clinicians.

Main Results:

• Successfully generated suboptimal radiotherapy dose distributions with statistically significant decreases in organ-at-risk sparing and target conformity, and significant increases in target hotspots (p < 0.05).
• The magnitude of dose alteration was controllable, allowing for tailored educational scenarios.
• Experienced clinicians rated the generated suboptimal dose distributions as realistic.

Conclusions:

• Developed and validated techniques for generating realistic, suboptimal radiotherapy dose distributions.
• These techniques enable direct manipulation of existing dose distributions without requiring a treatment planning system.
• The generated plans are perceived as realistic by experienced clinicians, offering a valuable tool for resident education.