Click here to

Session: Treatment Planning: Dose Calculation and Prediction [Return to Session]

BEST IN PHYSICS (THERAPY): Clinical Implementation of AI-Based 3D Dose Prediction Model as Physician Directive Support Tool to Streamline Head and Neck Planning Process

M Mashayekhi*, R McBeth, Z Trivedi, D Nguyen, S Jiang, M Lin, UT Southwestern Medical Center, Dallas, TX


SU-E-TRACK 5-6 (Sunday, 7/25/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: There exist some inherent shortcomings with the current treatment planning practice—e.g. deliverable dose to OARs is unknown a priori. In addition, the heterogeneity of clinical practices can lead to variation in the achievability of the planning goals and the resulted plan quality. The goal of this study is to incorporate an AI-based dose predictor into the treatment planning pipeline to develop a hybrid approach of physician intelligence and artificial intelligence that improves the planning efficiency, uniformity, and quality.

Methods: We developed a 3D AI-based dose predictor based on a hierarchically densely connected U-net model. The model is trained separately on definitive and post-operative plans. We integrated the model with Eclipse treatment planning system via Application Programming Interface (API) to allow seamless 3D dose and the uncertainty estimation in TPS upon the completion of contour. In phase one implementation, we tested the API and the expert model on 60 retrospective cases treated by 3 different physicians to build the confidence level of prediction quality for different practice styles. In phase two implementation, all four head and neck physicians use the dose predictor as a decision support tool to construct their planning directives for 60 patients prospectively. The physician directives and the achieved plans for phase one and two were compared to quantify the efficacy of AI-guided physician directive workflow.

Results: Our analysis shows increased practice uniformity among the physicians and improved achievability and plan quality of the plans upon application of AI. A dose difference of 3Gy was considered clinically significant. The frequency that resulted plan achieved within 3Gy from the physician directives has increased from only 47.5% to 86.7% upon implementing AI.

Conclusion: AI guided physician decision support improved the uniformity of practice, directive achievability, and the final plan metrics in a significant percentage of patients.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email