Click here to

Session: Therapy: External Beam: Automatic Treatment Planning [Return to Session]

Capturing Head-And-Neck Planning Trends with a Knowledge-Based Tradeoff Model

J Zhang1*, Y Sheng2, Y Ge3, Z Tian1, X Yang1, T Liu1, J Wu2, (1) Emory University, Atlanta, GA, (2) Duke University, Durham, NC, (3) University of North Carolina at Charlotte, Charlotte, NC


MO-IePD-TRACK 5-2 (Monday, 7/26/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: Since the clinical implementation of inverse planning, treatment planning knowledge has developed substantially, yet the knowledge shift is not well-understood. This study’s purpose is to analyze the progression of the clinical planning knowledge over time and investigate if knowledge-based planning (KBP) models built using historical data can represent the current planning practice.

Methods: To evaluate how clinical knowledge has progressed, we retrieved 304 anonymized head-and-neck treatment plans and binned them into four cohorts based on treatment date: 2011-2012, 2013-2014, 2015-2016, and 2017. Planning parameters and anatomic features were extracted from the treatment planning system via scripting. The first three cohorts were used for model training and the 2017 dataset was reserved for evaluation. A recently-proposed knowledge-based tradeoff model was used to quantify planning tradeoff considerations. Tradeoff directions projected on the 2017 dataset were recorded. Model prediction errors after taking tradeoff considerations were also evaluated. Additional plan information, including optimization constraints and optimization structures were analyzed.

Results: During the four time periods, the numbers of beams, structures with optimization constraints, and optimization constraints have not changed significantly (p>0.05). However, the standard deviations of these parameters have all decreased, indicating the planning practice has become increasingly standardized. Using the latest 2015-2016 dataset for model training reduces validation error by 26.0% for parotid, 21.9% for oral cavity, and 25.0% for pharynx, compared with models trained using cases from previous periods. However, the tradeoff decisions made clinically remain consistent: parotid and oral cavity have been involved in most tradeoff considerations.

Conclusion: We have demonstrated that the planning trends can be captured by a knowledge-based tradeoff model. Our results suggest that clinical planning knowledge changes over time and care must be taken when using prior treatment plans to guide new planning efforts.



    Treatment Planning, Modeling


    TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

    Contact Email