Statistical Motion Atlases for Standardized Heart Function Analysis across Multi-Modal Imaging

The heart works as a muscular pump, which needs a healthy amount of shortening and lengthening of heart muscle in each region of the heart to maintain good overall pump function. Diseases such as heart attacks, heart failure and heart rhythm disorders, as well as heart damage caused by cancer therapy drugs, are diagnosed and monitored by measuring pump function, including changes in regional heart contraction and motion. Strain is an engineering quantity which measures contraction and relaxation as a relative change in length. Accurate measurement of strain in all regions of the heart is vital to understand mechanisms of disease. Medical imaging methods such as echocardiography and cardiac magnetic resonance imaging are being used to measure regional strain, to help diagnose disease and monitor treatment, and to develop and evaluate computational analyses of heart function. However, current quantification methods are inaccurate and imprecise, since strain is highly sensitive to image artefacts, noise, and low resolution. Worse, strain estimates vary systematically between different imaging modalities and even between different commercial software products. Vendors use "black box" closed source solutions which hamper reproducibility. This leads to different standards and measures being used by different doctors. Subsequently, it is not known which of the many measures available is best for diagnosing heart disease and predicting outcomes. Clearly, a new way of solving this problem is required. This project will develop novel technologies for measuring motion and strain in the heart which are standardized between imaging modalities. We will use "artificial intelligence" neural network methods to automatically process different types of medical imaging examinations to obtain more accurate and precise strain measurements. These networks will be trained to learn how to predict the underlying motion and strain from thousands of image simulations, as well as thousands of patient scans, using a statistical atlas of heart motions. By simulating realistic images with exact high resolution heart motions derived from the statistical atlas, the networks will learn how to handle image artefacts, noise and low resolution in real images. More fundamentally, we will examine how statistical atlasing methods can help us discover which strain measures are best for diagnosing and predicting heart disease. We will then deploy these methods in high-throughput heart imaging clinics at St Thomas' Hospital, Royal Brompton Hospital, and other NHS hospitals. By making open-source tools widely available for doctors, we will test how standardised measurements and reports will work in practice. This will also reduce the costs of patient evaluation, by getting the information we need from commonly-performed scans and avoiding the need for specialized equipment. More accurate and precise evaluation of patients with heart disease will improve patient care, by identifying high risk patients and optimizing treatment dose. In particular, many cancer patients suffer from heart muscle damage caused by their cancer drug therapy. Our tools will enable better identification of which patients are at most risk and may require a change of treatment.