ESTRO 2022 - Abstract Book

S781

Abstract book

ESTRO 2022

Conclusion This novel hybrid deep learning tool has shown high accuracy and consistency in a validation set of lung cancer cases with challenging anatomical and imaging variations. By modelling dose uncertainty, this has promising applications in substructure dose calculations, radiotherapy plan optimisation, and feasibility of large cohort studies. Standardisation of use could also prove valuable for future studies on long-term cardiac toxicity.

MO-0890 Machine-learning based treatment couch parameter prediction for surface guided radiotherapy

G. De Kerf 1 , M. Claessens 2 , I. Mollaert 1 , W. Vingerhoed 1 , A. Sprangers 1 , D. Verellen 1,3

1 Iridium Netwerk, Radiotherapy, Wilrijk, Belgium; 2 University Antwerp, Medicine and Health Sciences, Antwerp, Belgium; 3 University of Antwerp, Medicine and Health Sciences, Antwerp, Belgium Purpose or Objective Recently, Surface Guided Radiation Therapy (SGRT) paves the way towards a complete replacement of patient’s tattooing with a markerless patient’s workflow as being accurate and reproducible [1] [2]. As early adopters we implemented SGRT on all our machines (10 linacs and over 6000 patients per year), raising the need of an automated and efficient workflow. Accuracy of SGRT-based initial patient’s setup hardly depends on curvature type and symmetry of patient’s body in the irradiated area and can be improved by using additional information like tattoos, fiducial markers or predicted couch parameters [3] [4]. This study aims to improve the SGRT workflow by predicting couch parameters via a machine learning approach to give RTTs additional guidance during initial patient setup on where patient is expected according to the TPS.

Materials and Methods Barium markers are placed underneath both department’s CT scanners couches (Philips and Siemens). This indexation makes the couch parameter prediction independent of any immobilization device, except for the Encompass SRS Immobilization System (Qfix) which floats beyond the couch top. The latter has embedded radio-opaque markers that will be used as reference point. On a CT image, marker detection starts with a rough estimate of the expected marker position in X and Y direction, followed by a thresholding step. Afterwards, a K-means clustering (n_clusters = 2) algorithm tries to detect the couch markers or the cranial Encompass reference point (n_clusters = 1). Finally, a post-processing step validates the detected markers and the expected treatment couch values are calculated (Figure 1). At first fraction, patients are positioned using an SGRT system (AlignRT, cRAD) before acquiring an image for final patient position verification. The treatment couch positions are captured at the moment of image acquisition to resemble the parameters that minimizes the SGRT spatial positioning deviations. For verification, the predicted couch coordinates of 99 treatments (29 SRS, 70 couch markers for lung, prostate or oligo meta treatments) are compared against the acquired parameters after the patient was positioned according to the SGRT instructions. Results Based on preliminary data, couch parameters could be predicted with an accuracy of 0.2mm ± 6.5 (see Figure 2). Highest accuracies were obtained for patient positioned with the Encompass system (0.2mm ± 1.7). Lowest accuracy is obtained for SBRT Lung patients (0.4mm ± 8.8) positioned on a thorax support and most variation was seen in the lateral direction.