ESTRO 38 Abstract book

S557 ESTRO 38

segmentation based only on CT images performed similarly to segmentation using PET images. Conclusion All segmentation approaches based on PET images performed satisfactorily due to the high SUV of the tumour relative to other tissues. However, only the deep learning approach was able to discern the tumour in CT images, without the need of extensive feature engineering as required by the linear classifier.

References 1. Ardenfors O, et al. Organ doses from a proton gantry- mounted cone-beam computed tomography system characterized with MCNP6 and GATE. Phys Med. 2018;53:56-61 2. de las Heras Gala H, et al. Quality control in cone-beam computed tomography (CBCT) EFOMP-ESTRO-IAEA protocol. Phys Med. 2017;39:67-72. PO-1009 Comparison of automatic tumour segmentation approaches for head and neck cancers in PET/CT images A. Rosvoll Groendahl 1 , M. Mulstad 1 , Y. Mardal Moe 1 , I. Skjei Knudtsen 2 , T. Torheim 3 , O. Tomic 1 , U.G. Indahl 1 , E. Malinen 4 , E. Dale 5 , C.M. Futsaether 1 1 Norwegian University of Life Sciences NMBU, Faculty of Science and Technology, Ås, Norway ; 2 Oslo University Hospital, Department of Medical Physics, Oslo, Norway ; 3 University of Oslo / Oslo University Hospital, Institute for Cancer Genetics and Informatics, Oslo, Norway ; 4 University of Oslo / Oslo University Hospital, Department of Physics / Department of Medical Physics, Oslo, Norway ; 5 Oslo University Hospital, Department of Oncology, Oslo, Norway Purpose or Objective The objective of this study was to assess methods based on PET thresholding, machine learning using a linear classifier, and deep learning for automatic tumour segmentation of head and neck cancers in PET/CT images. Material and Methods This retrospective study examines 197 head and neck cancer patients who underwent a combined 18F-FDG- PET/CT scan prior to radiotherapy. Three tumour segmentation approaches with different levels of complexity were compared: (i) PET thresholding techniques, (ii) voxelwise classification using the machine learning algorithm Linear Discriminant Analysis (LDA), and (iii) deep learning using convolutional neural networks (CNNs). Manual gross tumour delineations made by an oncologist and a nuclear medicine specialist were considered as the ground truth. PET thresholding was conducted using either absolute thresholding or a percentage of maximum SUV, optimized using the Dice similarity coefficient (DSC) or the Receiver Operating Characteristics (ROC). The LDA approach used input features engineered using voxel intensities and voxel neighbourhood information extracted from the original PET/CT images and images transformed using CT windowing, as well as point and two-dimensional transformations to highlight different image characteristics. Training data was down-sampled to 50/50 class-balance prior to training the LDA model. The deep learning approach used a U-net architecture with a Dice loss function. Weights were initialised using the He initialisation scheme and optimised using Adam with a 10 - 4 step size and default hyperparameters. Raw PET images and raw or windowed CT images without down-sampling were used for this approach. The segmentation performance on validation data was assessed using the Dice similarity coefficient, sensitivity, specificity and the ROC-AUC (Receiver Operating Characteristics–Area Under the Curve). Results For PET thresholding, optimizing the SUV threshold with respect to ROC rather than DSC resulted in somewhat better segmentation with higher and balanced sensitivity (0.83-0.87) and specificity (0.85-0.89) (Table 1). LDA segmentation based only on CT images performed poorly (DSC 0.17), whereas segmentation based on PET images gave high sensitivity, specificity, AUC and DSC (0.50). Thus, features engineered from CT images did not provide sufficient information for tumour segmentation using a linear classifier. CNN segmentation had the overall highest DSC (0.62) (Table 1, Figure 1). Interestingly, CNN

PO-1010 Clinical evaluation of deep learning delineation of head and neck OARs. W. Van Rooij 1 , H. Ribeiro Brandao 1 , A. Delaney 1 , B. Slotman 1 , W. Verbakel 1 , M. Dahele 1 1 VUMC, Radiotherapy, Amsterdam, The Netherlands Purpose or Objective Target and organ-at-risk (OAR) delineation is a key step in radiotherapy treatment planning and adaptive radiotherapy, but it is often time consuming and resource intensive. Additionally, it is subject to interobserver variability and may require considerable anatomical knowledge. Most available automated methods for delineation do not, in general, perform as well as desired, and/or are slow. Deep learning for automated delineation is promising, but its evaluation has typically only been presented in terms of similarity coefficients or observer ratings for a limited number of OARs. Therefore, this retrospective study evaluates (1) the geometric accuracy, and (2) the dosimetric (clinical) impact of using deep learning delineation on many OARs for head and neck cancer (HNC). Material and Methods 142 anonymized clinical datasets, consisting of clinically contoured CT scans and structure sets, were used to train a convolutional neural network (U-Net) and 15 were used as a test set. The objective was to automate the delineation of the following OARs: left and right submandibular gland, left and right parotid gland, larynx, cricopharynx, pharyngeal constrictor muscle (PCM), upper esophageal sphincter (UES), brain stem, oral cavity and esophagus. Each OAR was delineated both manually (MD,