ESTRO 2023 - Abstract Book

S261

Saturday 13 May

ESTRO 2023

Conclusion We compared the performance of two state-of-the-art DL algorithms, which have reported outstanding semantic segmentation results, when trained with a partially-labeled clinical database. The predicted contours were very accurate with both models for almost all OAR. The underperformance on three structures was probably driven by the presence of large tumors and other external devices, such as gastro and nasopharyngeal tubes, deforming the anatomy and hindering segmentation (Figure 2). Overall, nnU-Net showed better results in terms of accuracy and computational requirements. PD-0327 Clinical benefits of multi-modality gross tumor volume auto-delineation in head and neck cancer H. Bollen 1 , S. Willems 2 , F. Maes 3 , S. Nuyts 1 1 University Hospitals Leuven, Radiation Oncology, Leuven, Belgium; 2 Catholic University Leuven, Processing Speech and Image, Leuven, Belgium; 3 Catholic University Leuven, Processing Speech and Images , Leuven, Belgium Purpose or Objective Gross tumor volume (GTV) delineation for head and neck cancer (HNC) radiation therapy planning is time consuming and prone to interobserver variability (IOV). The aim of this study was (1) to develop an automated GTV delineation approach of primary tumor (GTVp) and adenopathies (GTVn) based on a 3D convolutional neural network (CNN) exploiting multi- modality imaging input from CT, PET and MRI (instead of planning CT only) as required in clinical practice, and (2) to validate its accuracy, efficiency and IOV compared to manual delineation in a clinical setting. Materials and Methods Two datasets were retrospectively collected from clinical cases of HNC patients containing planning CT and additional PET/CT (76 patients) or MRI (74 patients). CNNs were trained for GTV delineation with consensus delineation of two experienced radiation oncologists as ground truth, with either single (CT) or co-registered multi-modal (CT+PET or CT+MRI) imaging data as input. For validation, GTVs were delineated in on 20 new cases by two observers, once manually, once by correcting the delineations generated by the CNN. Accuracy and IOV were assessed by the volume overlap between different delineations and efficiency by the gain in delineation time. Volume overlap between delineations was assessed calculating the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD) and Mean Surface Distance (MSD). Results Both multi-modality CNNs performed better than the single-modality CNN and were selected for clinical validation. For the multimodal approaches, performance of the late fusion strategy was significantly better than early fusion. Mean DSC between the ground truth and predictions for GTVp and GTVn, respectively, was 77% and 79% for the CT+PET-LF CNN and 61% and 70% for the CT+T1Gd-LF CNN. Agreement between predicted and corrected delineations was 69% and 79% for CT+PET-LF CNN and 59% and 72% for the CT+T1Gd-LF CNN (Table 1). IOV decreased significantly (GTVp: 76% vs. 95%, GTVn: 86% vs. 96%). Time efficiency increased with 48% (8 vs. 15.5 min, p < 10-5). Examples of predictions are shown in Figure 1 with the ground truth in red and automated uncorrected delineations in green.