ESTRO 2023 - Abstract Book

S1362

Digital Posters

ESTRO 2023

Bladder Femur L Femur R Pelvis LN CTV Rectum Bowel

Median Time EC (mins) 4.4 Median Time AIC (mins) 0.5

2.1 0.8

2.1 0.7

20.0

11.6 5.4 4.1 1.0

12.0

5.0

3.0

The median DSC for all OARs was > 0.8. CTV scored 0.88 and LN, 0.71.

Conclusion Our results have shown that editing a Limbus auto-contour to a clinical standard was more time efficient than manually contouring structures for prostate plans, with a significant time saving benefit. DSC values showed good agreement between AIC and EC from the Limbus system. For all structures, substantial time was saved editing an AIC, when compared to generating a new EC. Our study shows that AI can safely be used as a substantial time saver in the prostate radiotherapy planning process. Larger studies are required to confirm these preliminary results.

PO-1660 Automated Clinical Treatment Planning for breast: from manual to auto planning in Clinical Practise

A. Vella 1 , A. Gallagher 1 , L. Stubbs 1 , R. Dodkins 1 , H. Singh 1 , S. Padmanaban 1

1 Oxford University Hospital, Medical Physics and Engineering, Oxford, United Kingdom

Purpose or Objective Breast radiotherapy typically employees wide tangential fields and forward/inverse planned segments to improve the homogeneity of the dose distribution within the breast. Inverse planned segments potentially result in the quality of the plans becoming more consistent between planners. However, inverse planning represents a more complex treatment planning technique for a site representing approximately one third of radiotherapy patients. Automated Clinical Treatment Planning (ACT) was conceived as a rapid and efficient tool to streamline breast radiotherapy treatment planning. It was developed using an in-house Eclipse Scripting Application Programming Interface (ESAPI) application to automate dose optimisation and efficiently produce breast high-quality Intensity-Modulated Radiation Therapy (IMRT) treatment plans. Materials and Methods Only requirement of ACT for Breast radiotherapy is of a pre-defined tangential field arrangement for breast PTV target volume. Automatic plans were generated starting from a simple automation protocol which consisted of the constraints for breast PTV and organs at risk (OARs) (lungs, heart). The performance of the automatic approach was evaluated in terms of treatment planning time, target coverage, target dose heterogeneity and OAR sparing. Plans for 20 test patients were evaluated and compared with manual Breast IMRT/FiF planning. Following a local audit on clinical patients, the initial release was improved (V1 to V2) to support planning with newly installed TrueBeams, latest Varian calculation algorithm, and tested on the same patient cohort. Results ACT-Breast was able to generate clinical acceptable treatment plans (single/mixed energy) in the evaluated patients. ACT Breast drastically reduced treatment planning times to ~15 minutes, with the actual ACT plan creation time ~2 mins, compared to ~1hr for manual planning. Target coverage was comparable at median D98 of 95.4% for ACT-Breast V1, 95.2% for V2 with AcurosXB algorithm respectively, against 95.5% for manual planning. Hotspots receiving V105% dropped from 3.2% for manual planning to 2.6% for ACT-Breast V1 and to 1% and 0.1% for V2/AAA and V2/AcurosXB respectively. Mean ipsilateral lung dose is comparable between manual and ACT-Breast V1, and slightly decreases for V2/AcurosXB with a variation between 0.1% and 0.2%. Mean heart dose in left-sided patients was ~0.5Gy on average for all the techniques, with a variation within 0.1Gy. Conclusion ACT-Breast automatically generates clinical suitable breast radiotherapy plans in a time efficient manner. ACT-breast offers a base for further improving dose constraints in challenging breast plans within a second optimisation run combining automated and manual planning where appropriate to maximise clinical care for patients. This technique offers to reduce the breast RT care path to <7 days and supports the per-protocol plan approval by RT Physics.

PO-1661 Deep learning based planning for clinical use following the CHHiP trial

R. Helander 1 , F. Löfman 1 , T. Atkins 2 , S. Whittle 2

1 RaySearch Laboratories, Machine Learning, Stockholm, Sweden; 2 Royal United Hospitals Bath, Medical Physics, Bath, United Kingdom Purpose or Objective Machine learning (ML) based treatment planning is a technique for automating the generation of deliverable treatment plans. This work investigates the process of developing and evaluating a deep learning (DL) based approach to plan generation following the hypofractionated arm of the CHHiP trial for prostate treatments [1]. The work highlights the importance of involving clinical experts during model development to achieve treatment plan quality that equals or supersedes clinical benchmark plans.

Materials and Methods