We use cookies to ensure that we give you the best experience on our website Learn more

May 16, 2019

ESOC-2019

AUTOMATED MULTI-FEATURE QUANTIFICATION OF PLAIN CT IN ACUTE STROKE

;

imaging

machine learning

artificial intelligence

ct

Abstract

Abstract

thumbnail

Keywords

imaging

machine learning

artificial intelligence

ct

Abstract

Background and Aims: Optimal decision-making in acute stroke depends upon evaluating multiple radiological features on unenhanced brain CT. While analytic software have recently been introduced to assist feature interpretation (e.g. “e-ASPECTS”), these have not been shown to distinguish multiple pathologies, that commonly co-occur. This is technologically challenging given that many relevant CT features are hypoattenuating, and overlap in appearance. Here we evaluate a novel, fully-automated method for separately quantifying infarction vs. leukoaraiosis vs. CSF on plain CT. Methods: Experts annotated plain CT heads for CSF (separately for cortex/ventricles/brainstem), leukaraiosis and established infarctions, in 50, 100 and 450 cases respectively. CTs were fine-cut (0.4 x 0.4 x 0.8mm). Annotations fed into two feature-specific, convoluted neural networks, and a random-forest model (the latter for leukoaraiosis), before being combined in a single 2.5D U-net. Validation was conducted in a separate set of 110 acute ischemic stroke cases. Tests included correlations of feature volumes, and spatial similarity (Dice score). Results: Imaging processing failure occurred in 5/110 cases. Correlations between automated and expert volumes for CSF, leukoaraoisis and infarcts were: r2: 0.92 , 0.71 and 0.82 (all p<0.001). Dice scores were respectively, 0.88, 0.65 and 0.63. However, for infarct volumes of >1cc, Dice score was 0.82. This represents a >10-30% improvement in segmentation accuracy compared to current methods for fully-automated CT infarction segmentation. Conclusions: Unlike existing automated stroke-imaging software, that focus on single features, we demonstrate a method allowing for accurate distinction and quantification of multiple hypoattenuating CT features in acute stroke.

Discover over 20,000 new abstracts, posters and presentations from leading academic conferences every month. Stay on top of the latest findings, methodologies and discussions happening in your research field around the world.

Company

Legal

Follow us

© Copyright 2019 Morressier GmbH. All rights reserved.

© Copyright 2019 Morressier GmbH.
All rights reserved.