< Back to previous page

Project

Deep learning for computer-aided diagnosis based on three-dimensional echocardiographic imaging

Echocardiography is the modality of choice when diagnosing cardiovascular diseases (CVD) not only due to the fact that it is non-invasive, mobile (even portable) and relatively cheap but also because it operates in real-time at a good temporal resolution. Despite its clinical popularity, interpretation of the images / extracted parameters can be challenging; even for experts. Thanks to the developments of artificial intelligence and machine learning – which allow modeling complex data patterns – this issue can be addressed by designing a computer-aided diagnostic (CAD) system. The overall goal of the current project is to develop, validate and clinically test new learning strategies towards computer-aided interpretation of three-dimensional (3D) echocardiographic data. The main questions of this project are whether implementation of a CAD system using echocardiographic data via ‘multi-task deep learning’ is feasible and if the deep learning strategy can bring added value over the available mechanistic approaches. This project has the three main goals. Developing a 3D segmentation and tracking (i.e. temporal registration) algorithm using ‘deep learning’ techniques in order to extract left ventricular (LV) morphological parameters as well as measure myocardial functional characteristics. Contrasting the above-developed algorithms against in-house available software tools for LV segmentation and motion estimation based on ‘mechanistic’ approaches using validated experimental as well as clinical databases and modeling the measured morphological and functional characteristics using the advanced statistical learning techniques and combining the modeled features in order to build a CAD system. In order to reach the goals set out in this project, we will retrospectively analyze a large database of 3D echocardiographic images that is available at the ‘Lab on Cardiovascular Imaging and Dynamics’ at the University of Leuven. In a first phase, a deep learning algorithm will be developed for segmentation of the echocardiographic images by building upon the state of the art algorithms. In a second phase, a mechanistic 3D segmentation algorithm which has been developed in our Lab based on B-spline explicit active surfaces (BEAS) will be employed to automatically segment myocardial borders from the same volumetric data sets. In a third phase, the morphological and functional parameters computed by deep learning will be evaluated by contrasting them with the available myocardial segmentation and motion ground-truth. Finally, a combined shape-function analysis will be done in order to correct local/global functional characteristics for local/global loading conditions and to simultaneously analyze morphological and mechanical characteristics of the heart. The obtained parameters will then be used in a classification framework for predicting some major adverse cardiac events (e.g. hospitalization due to heart failure or myocardial infarction, elective or urgent coronary artery bypass grafting) in patients in the DOPPLER-CIP database.

Date:9 Sep 2019 →  22 Sep 2023
Keywords:Deep learning, computer-aided diagnostic (CAD), Segmentation, Echocardiographic images, Modeling
Disciplines:Biomedical image processing
Project type:PhD project