Machine Learning for Automated Heart Strain and Motion from DENSE

点击翻译按钮获取中文摘要

Aim of the PhD Project:DENSE MRI provides accurate and reproducible myocardial strain, but processing currently relies on extensive manual input.This project will utilize existing data and state-of-the-art machine-learning algorithms to automatically process and segment DENSE data, outputting cardiac geometry as well as pixelwise displacement and strain.Project Description / Background:Recent advances in machine learning and artificial intelligence methods have enabled the development of new tools for the quantitative analysis of cardiac performance in medical imaging examinations. However, applications to patients require implementation and evaluation on clinical scans in specific disease cases. This project will develop new methods for the analysis of DENSE cardiac MRI exams. The resulting software tools will be used in clinical studies performed at the Royal Brompton Hospital.DENSE is the most accurate and best resolution method for non-invasive quantification of strain and motion of heart muscle [1,2]. However, the images currently require complex and time-consuming off-line post-processing to estimate the clinically important strain parameters. In particular, the borders of the heart muscle must be determined (segmentation) and the phase signal unwrapped (due to aliasing). In the presence of noise and artefacts this is difficult and errors need to be corrected manually. The post-acquisition nature of this processing also means that MRI technologists are operating "blind" to a large extent, with little indication of the quality of the strain data and results, which could be used to guide subsequent data acquisition. Recently, advances in machine learning and AI methods have shown promise in automatic evaluation of cardiac MRI data [3, 4]. These do not require manual interaction and can considerably speed up the evaluation process.This project will leverage recent advances in machine learning and artificial intelligence to automatically analyse DENSE data, including segmentation and phase unwrapping, to provide accurate pixel-wise strain information in all regions of the heart. Training will make use of the high performance computing clusters at Imperial and/or KCL. The trained models will be incorporated into online image reconstruction tools, providing immediate measures of myocardial displacement and strain at the scanner.