Aim 1: Develop a fully automated method for echo quantification and disease detection, focused on diseases of abnormal heart thickness. Automation of image analysis can be invaluable in both research and clinical contexts, by 1) increasing throughput to allow analysis of thousands of studies; 2) standardizing quantitation; 3) producing features for downstream statistical learning; and 4) providing lower cost diagnostic solutions when resources are limited. We will use three machine learning algorithms to achieve this Aim – convolutional neural networks, active appearance models, and Bayesian networks – all of which have been successfully used by the computer vision community to carry out comparable tasks, but have yet to be assembled for this clinical purpose. We will develop and apply an automated workflow to diseases of thickened hearts as follows: 1a. Develop an end-to-end workflow for automated echo quantification 1b. Develop an automated classifier to discriminate between three forms of cardiac thickening: CA, HCM and hypertensive heart disease. 1c. Develop and validate a fully automated approach for detection of CA and HCM. Aim 2: Characterize quantifiable measures of disease progression in HCM and CA and associate these with clinical outcomes. HCM and CA patients are routinely followed by serial echocardiography to guide clinical decisions. Nonetheless, given the sheer number of frames to consider, it can be challenging for echo readers to make comparisons across studies and detect any change. We hypothesize that applying computer vision analytic approaches to echo video loops in a semi-automated manner will allow extraction of informative features that can be compared across studies and used to detect early signs of cardiac dysfunction. Such changes will then be evaluated for association with clinical progression, including arrhythmias and symptom development. Through successful completion of these Aims we anticipate creating an innovative workflow for early diagnosis of cardiomyopathic disease as well as detection of disease progression. We also anticipate discovering novel informative features characterizing the changes in cardiac structure and function that accompany these specific diseases. Our ultimate goal in carrying out this work is to build a set of validated and teachable computational approaches that can be used by the greater cardiology research and clinical community. Such analytic approaches should also be widely applicable to other problems in cardiac disease, including monitoring cardiotoxicity of anti-neoplastic agents, and quantifying improvement in response to novel therapies.
|Effective start/end date||8/1/19 → 6/30/22|
- Brigham and Women's Hospital (120816//5R01HL140731-03)
- National Heart, Lung, and Blood Institute (120816//5R01HL140731-03)
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.