Deep Learning for Doppler Echocardiography from Limited Labeled Data

Abstract

Senior thesis:

Aortic Stenosis (AS) is a common cardiovascular disease that deteriorates the cardiac valves by causing the valve to narrow, reducing or blocking blood flow to the heart. When treated this disease has a low mortality rate; the problem is that it’s difficult to diagnose in its early stages. Thus, the Hughes Lab, in collaboration with cardiologist Dr. Benjamin Wessler, set out to diagnose AS using Deep Learning. Echocardiograms, the gold standard for diagnosing AS, are used to capture many different types of images of the heart. Past work has mostly focused on 2D structural images of the heart. However, this thesis mostly focused on a different type of echocardiogram: spectral Dopplers. Spectral dopplers access information about the velocity of blood through valves and provide vital information to cardiologists to help inform their diagnosis of valve conditions.

The first part of this thesis kept the focus on 2D structural images. In this small sub-project, to enable the assessment of cross-hospital generalization potential, we built a standardized benchmark task for view type classification using open- access datasets from 3 countries. A published paper (Huang et al., 2023) shows how new SSL methods can be validated on this benchmark task.

The rest of this project focused on creating a diagnosis classifier for spectral Dopplers. To do this, we first created three view classifiers using standard supervised learning and Positive Unlabeled Learning, to identify the relevant view types, and two diagnosis classifiers using standard supervised learning. Then we combined the results of these classifiers and report our performance on held-out patients and compare them to several possible baselines.