Automated assessment of echocardiographic image quality using deep convolutional neural networks

Labs, Robert (2022) Automated assessment of echocardiographic image quality using deep convolutional neural networks. Doctoral thesis, University of West London.

[thumbnail of Labs - Final PhD Thesis (April 22).pdf]
Labs - Final PhD Thesis (April 22).pdf - Accepted Version

Download (5MB) | Preview


Myocardial ischemia tops the list of causes of death around the globe, but its diagnosis and early detection thrives on clinical echocardiography. Although echocardiography presents a huge advantage of a non-intrusive, low-cost point of care diagnosis, its image quality is inherently subjective with strong dependence on operators’ experience level and acquisition skill. In some countries, echo specialists are mandated to supplementary years of training to achieve ‘gold standard’ free-hand acquisition skill without which exacerbates the reliability of echocardiogram and increases possibility for misdiagnosis. These drawbacks pose significant challenges to adopting echocardiography as authoritative modalities for cardiac diagnosis. However, the prevailing and currently adopted solution is to manually carry out quality evaluation where an echocardiography specialist visually inspects several acquired images to make clinical decisions of its perceived quality and prognosis. This is a lengthening process and laced with variability of opinion consequently affection diagnostic responses. The goal of the research is to provide a multi-discipline, state-of-the-art solution that allows objective quality assessment of echocardiogram and to guarantee the reliability of clinical quantification processes. Computer graphic processing unit simulations, medical imaging analysis and deep convolutional neural network models were employed to achieve this goal. From a finite pool of echocardiographic patient datasets, 1650 random samples of echocardiogram cine-loops from different patients with age ranges from 17 and 85 years, who had undergone echocardiography between 2010 and 2020 were evaluated. We defined a set of pathological and anatomical criteria of image quality by which apical-four and parasternal long axis frames can be evaluated with feasibility for real-time optimization. The selected samples were annotated for multivariate model developments and validation of predicted quality score per frame. The outcome presents a robust artificial intelligence algorithm that indicate frames’ quality rating, real-time visualisation of element of quality and updates quality optimization in real-time. A prediction errors of 0.052, 0.062, 0.069, 0.056 for visibility, clarity, depth-gain, and foreshortening attributes were achieved, respectively. The model achieved a combined error rate of 3.6% with average prediction speed of 4.24 ms per frame. The novel method established a superior approach to two-dimensional image quality estimation, assessment, and clinical adequacy on acquisition of echocardiogram prior to quantification and diagnosis of myocardial infarction.

Item Type: Thesis (Doctoral)
Subjects: Medicine and health
Depositing User: Robert Labs
Date Deposited: 07 Jun 2022 15:34
Last Modified: 07 Jun 2022 15:34


Downloads per month over past year

Actions (login required)

View Item View Item