Multiphase Computed Tomographic Angiography with Bone Subtraction Using 3D Multichannel Convolution Neural Networks

Multiphase computed tomographic angiography (CTA) have been demonstrated to be a reliable imaging tool for evaluating cerebral collateral circulation that can be used to select acute ischemic patients for recanalization therapy. We proposed using bone subtraction techniques to visualize multiphase CTA for clinicians to make fast and consistent decisions in the imaging triage of acute stroke patients. A total of 40 multiphase brain CTA datasets were collected and processed by two bone subtraction methods. The reference method used pre-contrast (phase 0) scans to create ground truth bone masks by thresholding. The tested method used only contrast enhanced (phases 1, 2, and 3) scans to extract bone masks with two versions (U-net and atrous) of 3D multichannel convolution neural networks (CNNs) in a supervised deep learning paradigm for semantic segmentation. Half (n = 20) of the datasets were used to train and half (n = 20) were used to test the conventional 3D U-net and a patch-based 3D multichannel atrous CNN. The tested U-net and atrous CNNs achieved a mean intersection over union (IoU) scores of 90.0% +/- 2.2 and 93.9% +/- 1.2 respectively.Clinical Relevance—This bone subtraction technique helps to visualize CTA volumetric datasets in the form of full brain angiogram-like images to assist the clinicians in the emergency department for evaluating acute ischemic stroke patients.