Click here to

Session: Deep Learning for Treatment Planning [Return to Session]

Input Feature Design and Their Impact On the Artificial Intelligence Planning Agent’s Performance

X Li1*, Y Ge2, Q Wu1, C Wang1, Y Sheng1, W Wang1, H Stephens1, F Yin1, QJ Wu1, (1) Duke University Medical Center, Durham, NC, (2) University of North Carolina at Charlotte, Charlotte, NC

Presentations

TU-I345-IePD-F2-5 (Tuesday, 7/12/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: Deep Learning (DL)-based IMRT planning through fluence map prediction could save human time and effort by avoiding lengthy inverse planning. Although DL-based planning has demonstrated promising preliminary outcomes, the ad hoc or black-box nature of DL models might undermine future development for more robust performances or generalizations. This study aims to examine how input feature design may influence the DL prediction performances.

Methods: This study included 231 HN IMRT patients, where 200/16/15 cases were assigned to training/validation/test groups. Ground truth (GT) plans were generated in batch using script to unify machine settings and delivery techniques. Three input feature designs of the DL model were investigated based on different hypotheses: 1) D1 assumes AI needs information of all critical structures from all beam angles to predict fluence maps; 2) D2 assumes that local anatomical information (within 3~5 cm patches) is sufficient for predicting the radiation intensity of a beamlet for the prospective beam angle; 3) D3 assumes the need for both local anatomical information and inter-beam interaction to predict radiation intensity values of the beamlets that intersect on a voxel. The outcomes of these 3 designs were analyzed using main dose-volume metrics. One-way ANOVA with multiple comparisons correction (Bonferroni method) was performed (significant level=0.05).

Results: For PTV-related metrics, AI plans in D1~3 had significantly higher maximum dose (p=2.8e-22), conformity index (p=3.1e-13), and heterogeneity index (p=1.6e-18) compared to GT plans, with D2 having the worst performance. Meanwhile, except for cord+5mm (p=2.6e-3), AI plans’ OAR dose was comparable to GT plans’.

Conclusion: Based on the experiment, local anatomical information contains most of the information that AI needs to predict fluence maps for clinically acceptable OAR sparing. However, information from all beam angles is necessary to make predictions with accurate PTVcoverages. These results could provide valuable insights for future DL models’ designs.

Funding Support, Disclosures, and Conflict of Interest: Varian medical systems (master research agreement)

Keywords

Radiation Therapy, Treatment Planning, Modeling

Taxonomy

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

Contact Email

Share: