AI of the beholder.

Treatment planning for head and neck cancer is complicated, and there can often be discrepancies between “safe” plans and optimal plans. We’ve seen several studies describing tools for predicting the optimal and achievable dose for organs at risk (OARs) prior to head and neck planning. In this study, the authors curated a set of 112 of the highest quality head and neck plans generated using RapidPlan knowledge based planning (KBP). They used these plans to train a deep learning model to predict the optimal dose distribution based on a patient’s CT scan, target volumes, and OARs. The model performed well in the testing phase in predicting the actual treatment planning generated by the KBP system. They went a step further, though, and evaluated the deep learning model’s ability to assess plan quality for manually generated clinical plans. Not only did it accurately predict the dose distribution of those 14 plans, it flagged 63 OARs for improvement (most frequently the mandible, brain, and cord). In comparison, a group of 3 physicians flagged 47 OARs for improvement. Surprisingly (or maybe not), there was striking variability among physician plan quality assessment. physician plan quality assessment was all over the place with 83% of the physician-flagged OARs only flagged by a single physician. The deep learning model identified the majority (64%) of OARs flagged by the physicians. | Gronberg, Pract Radiat Oncol 2023