Deep Learning-based Automated Delineation of Head and Neck Malignant Lesions from PET Images

Accurate delineation of the gross tumor volume (GTV) is critical for treatment planning in radiation oncology. This task is very challenging owing to the irregular and diverse shapes of malignant lesions. Manual delineation of the GTVs on PET images is not only time-consuming but also suffers from inter- and intra-observer variability. In this work, we developed deep learning-based approaches for automated GTV delineation on PET images of head and neck cancer patients. To this end, V-Net, a fully convolutional neural network for volumetric medical image segmentation, and HighResNet, a 20-layer residual convolutional neural network, were adopted. 18F-FDG-PET/CT images of 510 patients presenting with head and neck cancer on which manually defined (reference) GTVs were utilized for training, evaluation and testing of these algorithms. The input of these networks (in both training or evaluation phases) were 12×12×12 cm sub-volumes of PET images containing the whole volume of the tumors and the neighboring background radiotracer uptake. These networks were trained to generate a binary mask representing the GTV on the input PET subvolume. Standard segmentation metrics, including Dice similarity and precision were used for performance assessment of these algorithms. HighResNet achieved automated GTV delineation with a Dice index of 0.87±0.04 compared to 0.86±0.06 achieved by V-Net. Despite the close performance of these two approaches, HighResNet exhibited less variability among different subjects as reflected in the smaller standard deviation and significantly higher precision index (0.87±0.07 versus 0.80±0.10). Deep learning techniques, in particular HighResNet algorithm, exhibited promising performance for automated GTV delineation on head and neck PET images. Incorporation of anatomical/structural information, particularly MRI, may result in higher segmentation accuracy or less variability among the different subjects.