ESTRO 2023 - Abstract Book

S1310

Digital Posters

ESTRO 2023

We have developed AI models for whole breast and head-and-neck (HN) radiation therapy treatment planning. The breast AI model is built upon a random forest algorithm and predicts fluence maps of single and dual-energy tangential beams. The HN AI model is built upon a conditional generative adversarial network (cGAN) and predicts fluence maps of nine equally spaced beams. In addition, a graphical user interface (GUI) is developed to integrate the AI tools within the clinical treatment planning system (TPS), and also allows the AI algorithm to be executed on a designated workstation. The designated workstation is configured specifically for the AI algorithm with graphics processing unit (GPU) support. The AI plans are imported into the TPS for dose calculation, followed by an optional automatic fine-tuning and final physician approval (Final plan). Planners can visually examine the dose distribution and make further adjustments as deemed clinically necessary. Results The average plan generation time including user interactions was 6-15min/case, compared to 45 min to 2 hours of typical manual planning. The additional fine-tuning option provided not only a viable flexibility to customize an individual patient’s plan, but also ease of mind support. For the total of 1600 patients in 3 years of auto planning, the dosimetry between the AI plan and final plan (with human edit) showed no statistical difference, except for 105% hot spot in the PTV. Further, there is a steady decrease in the usage of the fine-tuning option, from 4.0% in 2019 to 0.8% in 2021, indicating increased acceptance and trust of AI planning.

*MMV: manual modification (of the fluorene) value

Conclusion The in-house AI planning tools have been successfully implemented in clinical environment and have demonstrates robust performance and significant proficiency improvement in the clinical workflow.

PO-1616 The importance of regional analysis in contour comparison

H. Chamberlin 1 , E. Henderson 1 , R. Cowan 2 , C. Anandadas 2 , F. Wilson 2 , C. Hague 2 , C. Chan 2 , L. Davies 3 , M. Daly 1 , C. Eccles 3 , M. Harris 2 , M. Maxwell 3 , Z. Oong 2 , Z. Turpin 3 , J. Radford 2 , S. Howell 2 , S. Astley 4 , M. Aznar 1 , E. Vasquez Osorio 1 1 University of Manchester, Radiotherapy Related Research Group, Division of Cancer Sciences, Manchester, United Kingdom; 2 The Christie NHS Foundation Trust, Department of Clinical Oncology, Manchester, United Kingdom; 3 The Christie NHS Foundation Trust, Radiotherapy, Manchester, United Kingdom; 4 University of Manchester, Division of Informatics, Imaging and Data Sciences, Manchester, United Kingdom Purpose or Objective Metrics such as the Dice coefficient (DSC) and mean distance to agreement (mDTA) are often used to quantify delineation variation between multiple observers or the performance of autocontouring, but fail to capture variations in specific regions where contours disagree. These variations have implications for margin definition and RT dose to the target and healthy tissues. Here, we demonstrate the value of regional analysis applied to variation quantification between multiple observers in regions of the breast. Materials and Methods CT scans of 10 female lymphoma patients (age 15 to 34) were selected, and a reference breast contour was drawn manually and peer-reviewed by 2 breast oncologists using ESTRO consensus guidelines. In addition, auto-contours were produced using an atlas-based contouring tool in RayStation (RaySearch, Sweden) using 10 templates (also to ESTRO guidelines) and