Echocardiographic image analyses by deep learning. An opportunity to improve chamber quantification

Abstract

In cardiology, the quantification of left heart chamber volumes is a key task for prognosis and treatment decisions. Echocardiography is the most widely used method for chamber quantification, but current methodologies are time-consuming and associated with considerable observer related variability. It has recently been proposed that deep learning based methods for automation of the measurement process may be beneficial but there is a need to evaluate the clinical utility of this new technology. This thesis consists of three studies. In Study I, we developed dedicated segmentation networks tailored for the task of automated measurements of left ventricular volumes and ejection fraction. The performance of the method was evaluated both in real time and in automatic analyses of stored echocardiograms. We showed that the real time implementation led to a notable reduction in total time for image acquisition and measurements. We also showed that the precision of automated measurements was comparable to the precision of repeated measurements performed by the same clinician and better than the precision of different observers. Further, we performed a rigorous assessment of agreement between automated measurements and manual measurements by experts and discussed how the variability of the reference measurements influenced our results. Study II describes the technical development of dedicated segmentation networks for measurements of left atrial (LA) volume in LA focused echocardiograms. We experimented with different network architectures and strategies for selecting the end-systolic frame of the echocardiograms and found that the best results were achieved using a lightweight U-Net with 2 million parameters and a dedicated convolutional neural network trained to detect the relevant frames of the ultrasound clip. Study III contains a clinical validation of the method developed in Study II. We evaluated the method for automated measurements of LA end-systolic volume in internal and external databases and finally in real time during echocardiographic scanning. We found that the automated measurements were well aligned with manual reference measurements from echocardiography. We also performed a comparison against reference measurements from cardiac magnetic resonance imaging (CMR) and showed that due to intrinsic differences between the modalities, CMR is not a suitable reference for validation of automated measurements from echocardiography. We also evaluated the method in a sample with severely dilated atria due to longstanding rheumatic heart disease and discussed the benefits of the explainability of our sequential measurement pipeline, mimicking the approach taken by cardiologists. Finally, we showed how the real time implementation clearly improved efficiency of the measurement process.