ESTRO 2021 Abstract Book

S658

ESTRO 2021

Conclusion We created a CNN model which can generate clinically acceptable plans for locally advanced, left sided breast cancer patients. This model shows great potential to speed up the treatment planning process and to maintain consistency in plan quality. [1] Dan Nguyen et al 2019 Phys. Med. Biol. 64 065020

Poster discussions: Poster discussion 17: New technologies in image guidance

PD-0824 Automatic validation of organ at risk delineations based on machine learning C. Brink 1,2 , U. Bernchou 1,2 , I. Hazell 1 , V.N. Hansen 1 , A. Bertelsen 1 , E.L. Lorenzen 1 , N. Gyldenkerne 3 , R. Zukauskaite 3 , J. Johansen 3 , C.R. Hansen 1,2 1 Laboratory of Radiation Physics, Department of Oncology, Odense University Hospital, Odense, Denmark; 2 Department of Clinical Research, University of Southern Denmark, Odense, Denmark; 3 Department of Oncology, Odense University Hospital, Odense, Denmark Purpose or Objective The quality of delineated organs at risk (OAR) directly impacts the quality of the final radiotherapy treatment plans. As many organs are delineated in modern radiotherapy, a safe, fast, and automated evaluation of these is essential – in particular, if daily re-delineation is required for online adaptation. The current study aimed to develop an automatic system that can detect unintended OAR variations based on a machine learning algorithm Materials and Methods Training and validation cohorts, each including 100 previously treated Head and Neck (H&N) patients, were included in the study. The available OARs were used without any modifications prior to analysis. The outlier detection algorithm supports any number of organs and uses those available for a given patient. The current study included 23 OARs that are typically used in the treatment planning of H&N patients. In total, 13 OAR features were calculated for each available delineations. The features included: position of centre of mass (CoM), volume, surface area, compactness, lengths and orientations of the OAR axes, and distance from CoM to the nearest surface point. A feature model library was made from the training cohort. All patients were compared with the library. The comparison is based on the OAR available for the given patient and used a principal component analysis of the related library features. For each OAR, the analysis calculates a shape- related p-value indicating the degree of deviation from the library. Furthermore, a combined p-value related to deviations of the relative positions of all the OARs was calculated. The outlier detection threshold was 10^- 6 Results In the training cohort, 14 patients were identified as having at least one OAR as an outlier; clinical inspection found 13 of these to be true deviations. In the validation cohort, 23 patients were identified as outliers, with 13 being relevant and 10 false-positive. Some of the variations detected were left-right confusion of lens, mandible offset from the correct position, organs partly delineated within another organ, brainstems delineated too short, and islands of structures outside the intended volume (see figure examples). The detected variations did not necessarily impact the clinical treatment plan since the OAR might be placed far from the irradiated region. However, in retrospective toxicity studies, it is essential to detect all these deviations