ESTRO 2020 Abstract book

S370 ESTRO 2020

reviewed (271cm 3 , IQR 179-393cm 3 ) contours (p<0.001). Larger differences were seen across the dose-normalised DVH difference curve for rectum (Figure 1.A) than bladder (Figure 1.B). Figure 1.A Rectum

carcinoma(HNSCC) and compare the performance of different combinations of multimodality imaging. Material and Methods We collected 153 oropharyngeal HNSCC patients with CT, PET, and MR(T1 and T2) scans in a retrospective cohort. All patients had a clinical delineation of the tumor (GTVt) and involved lymph nodes (GTVn) performed on the CT. The delineations were approved according to protocol in a collaboration between an oncologist, a radiologist and a nuclear medicine physician considering all imaging modalities, and was considered the ground truth. We randomly divided the patients into three groups, 92 for training, 31 for validation, and 30 for testing. For deep learning, we used the 3D Unet with mixed residual connections, focal loss + dice loss as cost function, and the ADAM optimizer. The GTVt and GTVn were combined to GTVs being our segmentation target for training. All imaging data were deformably registered to the planning CT, and sliced on an isotropic 1 mm grid. We use a patch size of 64*128*128, 64 patches per patient, 20% axial plane rotation and inversion, augmentation, and trained for 40 epochs. We trained the network three times, using different combinations of images: 1) CT-PET-MR; 2) CT-PET; 3) CT- MR. Dice similarity coefficient (DSC), Hausdorff Distance 95 percentile (HD95) and mean (HDmean) were used as evaluation metrics on the test set, and report population averages plus/minus one standard deviation. Results In 26/30 test set cases, we were able to segment a GTV overlapping the ground truth (Fig 1). There were four cases where GTVs was not segmented in any of the imaging combinations. These outliers are excluded from the following numbers. The best result was obtained with the CT-PET-MR combination: DSC= 0.74±0.13, HD95= 9.54±7.3mm, HDmean= 2.76±2.04mm. The result for CT- PET was comparable with DSC= 0.73±0.12, HD95= 8.76±7.42mm, HDmean= 2.66±2.16mm, whereas a poorer result was obtained with CT-MR (DSC= 0.58±0.18, HD95= 12.92±7.49mm, HDmean= 3.73±1.81mm) (Fig. 2). Even better results might be achieved by a screening of the optimal threshold for binarization of the output probability map or other post-processing methods.

Figure 1.B Bladder

Figure Caption Difference between original and reviewed OAR relative cumulative DVHs. Shown for rectum (A) and bladder (B). Dose is rescaled so prescription dose = 100%. Legend shows relation to dose variations. Thick black line = mean Central review of OAR contours significantly increased the number of treatment plans which failed relative cumulative rectal dose-volume constraints. Centres may benefit from more detailed contouring guidelines and ongoing QA during trial conduct, to ensure OAR consistency. difference. Conclusion PH-0603 Deep learning delineation of GTV for head and neck cancer with multi-modality imaging J. Ren 1 , J. Nijkamp 1 , J.G. Eriksen 2 , S.S. Korremann 1 1 Aarhus University Hospital, Department of Oncology, Aarhus, Denmark ; 2 Aarhus University Hospital, Department of Experimental Clinical Oncology, Aarhus, Denmark Purpose or Objective In radiation oncology, segmenting tumor and pathological lymph nodes is a time-consuming and uncertain step in treatment planning. We investigate automated delineation with deep learning in head and neck squamous cell