ESTRO 2023 - Abstract Book

S672

Monday 15 May 2023

ESTRO 2023

Conclusion Presenting the tumour probability map with adjustable probability thresholds as output from the DL tumour contouring model gives clinically useful additional information for the radiation oncologist that can be used to optimize the tumour contouring process. MO-0801 Comparison of AI-based autosegmentation techniques exploiting prior knowledge at a 0.35 T MR-Linac M. Kawula 1 , I. Hadi 1 , L. Nierer 1 , M. Vagni 2 , D. Cusumano 2 , L. Boldrini 2 , L. Placidi 2 , S. Corradini 1 , C. Belka 1,3 , G. Landry 4 , C. Kurz 1 1 University Hospital, LMU Munich, Radiation Oncology, Munich, Germany; 2 Fondazione Policlinico Universitario “Agostino Gemelli” IRCCS, -, Rome, Italy; 3 German Cancer Consortium (DKTK), -, Munich, Germany; 4 University Hospital, LMU Munich, , Radiation Oncology, Munich, Germany Purpose or Objective The benefits of adaptive radiation treatment at MR-Linacs come at the cost of tedious and repetitive expert (re)contouring. The aim of this work was to compare AI-based methods for the autosegmentation of organs at risk (OARs), i.e., rectum and bladder, exploiting prior knowledge available at the irradiation stage in MR-guided radiation therapy (MRgRT) of prostate cancer patients. Two approaches were tested: patient-specific (PS) U-Net models and networks predicting deformation vector fields (DVFN). Materials and Methods Patients from two institutes operating 0.35 T MR-Linacs were included in the study. Cohort 1 (C1), comprising 73 planning images with manual delineations, was used for training of the benchmark U-Net model (BM). The second cohort (C2) included 19 patients with expert-contoured MRIs. One planning MRI and 5-33 fraction MRIs were available for each C2 patient (19 planning, 240 fraction images). PS models were generated for each patient by fine-tuning the BM using a planning image and validating it with the corresponding fraction data. 10 C2 patients were selected randomly for the PS hyperparameter optimization. The DVFN architecture was based on a U-Net with a spatial transformer layer and aimed to predict deformation fields between the planning and fraction images. Training was performed on the same 10 C2 patients (10 planning and 120 fraction images). The loss function was based on the L2 image similarity and the multi-scale dice similarity coefficient (DSC) between structures of interest. All models shared the same test set, i.e., fraction data from the remaining 9 C2 patients. The models were evaluated with DSC, the average (HDavg) and the 95th percentile (HD95) Hausdorff distance. Results Table 1 shows the evaluation of PS and DVFN models and compares them to the BM. The PS network with DSC for bladder/rectum of 0.93/0.90 performs better than both the DVFN (0.79/0.77) and the BM (0.91/0.87). The same was observed for HDs. Figure 1 shows slices that illustrate the performance of the methods under consideration. The PS approach can correct pronounced BM mistakes and determine rectum ends properly due to the inclusion of planning knowledge. The