Prescription tradeoff decision support for pancreas stereotactic body radiation therapy: From templates to artificial intelligence models

Abstract

Objective

To develop an artificial intelligence (AI) driven plan tradeoff decision making assistant for radiation oncologists.

Methods

A user interface (UI) was developed to integrate the assistant with the treatment planning system to facilitate prescription decision process. The assistant is powered by a machine learning core which was trained to learn the balance between planning target volume (PTV) coverage and organs-at-risk (OAR) sparing. A group of 98 pancreatic stereotactic body radiation therapy (SBRT) cases were retrospectively included for this study. The clinical plan's PTV coverage was compared against the model predicted value. A 10-fold cross validation was performed for all cases. The comparison was further analyzed in detail for three attending physicians. Cases with large discrepancy were identified and analyzed, and a replan was created to evaluate the achievability of the prediction.

Results

The clinical plan PTV V_100% was (87.7 ± 14.5)% while the model predicted value was (90.5 ± 9.6)%. Model agreement discrepancy was observed between attending physicians. Among all 98 cases, 9 were identified with large variation from the model prediction. For the replans, an average of 15.3% improvement was achieved over the original clinical plan, while OARs constrains were met.

Conclusions

The assistant's decision provides decent plan quality guidance for prescription drafting. It could provide valuable input prior to treatment planning and save valuable dosimetrist team and radiation oncologist effort. It could further provide valuable insight for resident education and training.