Abstract:
Recent advancement of deep learning-based algorithms has greatly improved the field of medical image analysis and computer-aided diagnosis/prognosis. Convolutional Neural...Show MoreMetadata
Abstract:
Recent advancement of deep learning-based algorithms has greatly improved the field of medical image analysis and computer-aided diagnosis/prognosis. Convolutional Neural Network (CNN) has shown superior accuracy and generalizability in performing prediction/classification tasks, thanks to its good utilization of the grid-like structure of input images in Euclidean space. In practice, one of the challenges in using classical CNN is the multi-size nature of medical images, which is especially prominent when the input images are from specific target region of interest (ROI) (e.g. tumor). Image sizes of those ROIs can vary a lot across patients, making the images difficult to be analyzed by CNNs where constant-sized inputs are expected. To address this challenge, we propose the Deep Voxel-Graph Convolution Network (DVGCN). DVGCN represents input images as their affinity graph and performs graph convolution to extract discriminative features. It then utilizes a sortpooling layer to sort the nodes and unifies the feature size used for prediction across images, thus solves multi-size challenge without explicitly resizing images. DVGCN is tested on 3D Positron-Emission Tomography (PET) images to predict the patient's cancer staging, its performance is compared with classical 3DCNN (with image padding) and radiomics models.
Date of Conference: 08-11 April 2019
Date Added to IEEE Xplore: 11 July 2019
ISBN Information: