2019/08/22 by Antonio García-Uceda Juárez, Raghavendra Selvan, Juarez, Antonio Garcia-Uceda +5 · 1 citation
Medicine · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Head and Neck Cancer Studies #Image and Video Processing (eess.IV) #Lung Cancer Diagnosis and Treatment #Radiomics and Machine Learning in Medical Imaging #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1908.08588
openalex publication_date 2019/08/22 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
We present an end-to-end deep learning segmentation method by combining a 3D\nUNet architecture with a graph neural network (GNN) model. In this approach,\nthe convolutional layers at the deepest level of the UNet are replaced by a\nGNN-based module with a series of graph convolutions. The dense feature maps at\nthis level are transformed into a graph input to the GNN module. The\nincorporation of graph convolutions in the UNet provides nodes in the graph\nwith information that is based on node connectivity, in addition to the local\nfeatures learnt through the downsampled paths. This information can help\nimprove segmentation decisions. By stacking several graph convolution layers,\nthe nodes can access higher order neighbourhood information without substantial\nincrease in computational expense. We propose two types of node connectivity in\nthe graph adjacency: i) one predefined and based on a regular node\nneighbourhood, and ii) one dynamically computed during training and using the\nnearest neighbour nodes in the feature space. We have applied this method to\nthe task of segmenting the airway tree from chest CT scans. Experiments have\nbeen performed on 32 CTs from the Danish Lung Cancer Screening Trial dataset.\nWe evaluate the performance of the UNet-GNN models with two types of graph\nadjacency and compare it with the baseline UNet.\n