AI Evaluation of Heart Muscle Function

Cardiovascular disease is the most common cause of death worldwide; early detection of cardiac dysfunction enables clinicians to improve therapies and care for patients experiencing heart disease. Myocardial strain is a direct measure of heart muscle function, representing the relative deformation of the myocardium throughout the cardiac cycle. Strain has been shown to be clinically significant for the early detection, diagnosis and prognosis of heart disease.

Quantifying strain non-invasively requires the acquisition of sequential imagery throughout the cardiac cycle, which is a task well suited to cardiac magnetic resonance (CMR). However, the regional assessment of strain using standard CMR sequences is not accurate with current methods, due to a perceived lack of image features within the myocardium. To reliably quantify regional strain, specialized MR acquisitions can be made, by spatially modulating the tissue magnetization, resulting in grid lines (tags) to appear and deform with the tissue. However, in clinical practice, tagged images are not acquired, and so clinical adoption of myocardial strain is hindered.

Recently, neural networks have been extremely successful in medical imaging tasks. Existing work has shown that myocardial strain can be automatically derived from tagging images, using neural networks. To address our research problem, we will improve existing works by imposing constraints on the predicted deformations based on cardiac mechanics. By applying our developed network to a dataset with paired tagging/standard imagery, we will be able to quickly and automatically compute "ground truth" deformation to train a new network for standard CMR sequences. The developed methods will be evaluated against strain values derived from manual analysis of tagging images and validated by retrospectively predicting cardiac outcomes from a cohort of cancer therapy patients.