Addressing the Gap in Oncologist Training

Purpose: Radiation oncology residents report limited understanding and confidence in assessing radiotherapy plan quality, as a result of limited exposure to diverse cases and the high-stakes clinical environment of residency training. To address the need for training in plan quality assessment, we developed techniques to create realistic head-and-neck (HN) and gynecological (GYN) dose distributions that are suboptimal in a controllable way. These suboptimal plans can be used as case examples for resident training on plan quality assessment.

Methods: High-quality dose distributions were first generated with a pre-trained deep learning model (3D Dense Dilated U-Net model trained using only high-quality plans). An automated “geometry-aware convolution” technique was developed to create suboptimal dose distributions with (1) decreased organs-at risk sparing (10%-60% increase in Dmean) or (2) decreased target conformality (5%-10% decrease in V95%). An automated random walk-based technique was developed to (3) introduce hotspots into the target (106%-113% of prescription). In total, 2148 suboptimal HN dose distributions and 658 suboptimal GYN dose distributions were generated from 39 and 20 HN and GYN dose distributions, respectively. Experienced clinicians reviewed 135 randomly selected suboptimal examples to assess realism on a 5-point Likert scale.

Results: The algorithms successfully created radiotherapy dose distributions with decreased organs-at-risk sparing, decreased target conformality, and increased hotspots that were statistically significant (p<0.05) when assessed by dose-volume histogram metrics. The resulting dose distributions were scored as realistic (Likert score ≥3) by experienced clinicians: 89%/79% by physicians, 48%/71% by physicists, and 54%/85% by dosimetrists, for HN/GYN, respectively.

Conclusion: In this study, we developed techniques to generate realistic but suboptimal dose distributionwithout the need for a treatment planning system. The resulting dose distributions appear realistic to experienced clinicians, particularly physicians, and can be used as educational material to support resident training on plan review in a low-stakes environment.