ESTRO 2023 - Abstract Book

S155

Saturday 13 May

ESTRO 2023

Department of Radiology, Boston, USA; 20 Massachusetts General Hospital - Harvard Medical School, Department of Radiology, Boston, USA; 21 Dana-Farber Cancer Institute - Harvard Medical School, Department of Imaging - Nuclear Medicine, Boston, USA; 22 Medical Center - University of Freiburg, Department of Radiation Oncology, Freiburg, Germany; 23 German Cancer Consortium (DKTK), Partner Site Freiburg, Freiburg, Germany; 24 German Oncology Center - University Hospital of the European University, Department of Radiation Oncology, Limassol, Cyprus Purpose or Objective With novel radio therapeutic treatment approaches, like focal dose escalation, for primary prostate cancer (PCa) on the rise, an accurate delineation of gross tumor volume (GTV) in prostate-specific membrane antigen PET (PSMA-PET) becomes more and more important. Manual approaches to this task are time consuming, require a lot of clinical expertise and are observer dependent. To overcome this issue, a fully convolutional neural network (CNN) was trained to automate this task and to provide an efficient solution to this problem. Materials and Methods A modified 3D U-Net was trained on a dataset consisting of 128 different 18F-1007-PSMA-PET images from three different institutions (Freiburg: n=77, Cyprus: n=32, Munich: n=19). Testing was done on one independent internal cohort (Freiburg: n=19) and two independent external cohorts (Boston: n=9 18F-DCFPyL-PSMA, Dresden: n=14 18F-1007-PSMA). Expert contours were generated in consensus for each patient using a validated technique. CNN predictions were compared to expert contours using Dice similarity coefficient (DSC). Co-registered whole-mount histology was used for the internal testing cohort to compare sensitivity and specificity of CNN predictions and expert contours. Results Median DSCs were Freiburg: 0.82 (IQR: 0.73-0.88), Dresden: 0.71 (IQR: 0.53-0.75) and Boston: 0.80 (IQR: 0.64-0.83) respectively. CNN sensitivity was slightly higher than manual contours with a median sensitivity of 0.88 (IQR: 0.68-0.97) and 0.85 (IQR: 0.75-0.88), respectively. With a median specificity of 0.83 (IQR: 0.57-0.97) and 0.88 (IQR: 0.69-0.98) for CNN and expert contours, respectively, CNN specificity was slightly lower (p<0.05). Conclusion Deep learning methods provide a great way to automate complex tasks in modern medicine. Even partially achieving higher sensitivities than experts while not sacrificing too much specificity, fast and reliable GTV segmentation can be obtained for 18F-1007-PSMA-PET and 18F-DCFPyL-PSMA-PET.

Description: Evaluation process for specificity and sensitivity for GTVs on Freiburg testing cohort. PCa GTVs are delineated by both CNN and experts separately. Co-registered whole mount histology is used as ground truth.

Mini-Oral: Image processing and treatment evaluation

MO-0222 A neural network to create super-resolution MR from multiple 2D brain scans of paediatric patients J. Benitez-Aurioles 1 , A. Davey 2 , M. Aznar 2 , A. Bryce-Atkinson 2 , E.M. Vásquez Osorio 2 , S. Pan 3 , P. Sitch 3 , M. Van Herk 2 1 University of Manchester, Division of Informatics, Imaging and Data Sciences, Manchester, United Kingdom; 2 University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom; 3 The Christie NHS Foundation Trust, Department of Proton Therapy, Manchester, United Kingdom Purpose or Objective High-resolution (HR) MRI provides detailed soft-tissue information that is useful in assessing long-term side-effects after radiotherapy in childhood cancer survivors, such as facial asymmetry or morphological changes in brain structures. However, 3D HRMRI requires long acquisition times, so in practice often multiple 2D low-resolution (LR) images (with thick slices in multiple planes) are acquired for patient follow-up. In this work, we present a super-resolution (SR) convolutional neural network (CNN) which can reconstruct a HR 3D image from 2D LR images, in order to improve the extraction of structural biomarkers from routine scans. Materials and Methods A multi-level densely connected super-resolution CNN [1] was adapted to take two perpendicular LR scans (e.g., coronal and axial) as tensors and reconstruct a 3D HR image. Scans were resampled to a resolution of 1mm 3 before being fed into the network. A training set of 80 HR T1 paediatric (9-10years, healthy subjects) head scans from the Adolescent Brain Cognitive Development (ABCD) study was used as baseline, and 2D LR images were simulated to use as input into the CNN (Figure 1). 10 additional scans from ABCD were used to tune the hyperparameters of the CNN. The output of the model (images CNN ) was compared against simple interpolation (resampling and averaging both inputs), (images interp ). The evaluation was done in two steps. First, the quality of the reconstructed HR images was assessed using the peak signal to-noise ratio (PSNR) (larger values indicate better quality) compared to baseline. Secondly, the precision of structure