Abstract 17562: Automatic Quality Assessment of Echo Apical 4-chamber Images Using Computer Deep Learning
Background: Echocardiography is a noninvasive imaging modality that allows for assessment of cardiac structure and function for clinical decision-making. However, scanning is a resource intensive technical procedure, which is dependent on operator skill and experience. We sought to develop a method to automatically compute an image quality score of a specific echo image (end-systolic apical 4-chamber view) for future application to a real-time feedback platform for optimization of images at the time of acquisition.
Methods: The model used for this application was a deep convolutional neural network model trained on a large set of labeled samples. Randomly fetched end-systolic apical 4-chamber images were obtained from a clinical database. A total of 6 916 images were manually graded by a single observer for image quality with a score ranging between 0 (unacceptable image) and 5 (good quality). Of these labeled images, 80% were used to train the network and the remaining 20% were used to test the model for agreement with manual scoring.
Results: The performance of the model is based on the 1 387 images for testing. The average absolute error of the model compared with manual scoring was 0.68±0.58, with 91% of the images obtaining a score difference of <1. Using a sample of 200 images, intraobserver variability demonstrated high agreement; within subject standard deviation was 0.65 (κ = 0.80). The average time required for the network computation of the image quality score was 10 ms.
Conclusion: Deep convolutional neural networks can reproducibly and accurately score the quality of apical 4-chamber images as compared with manual scoring.
Future direction: With the rapid computation time, this study demonstrates that real-time echo image quality feedback is feasible. We plan to expand the network for application to other views across the cardiac cycle. Real time feedback of echo quality may motivate optimization of echo images, particularly for novice scanners and nonsonographers. This technology has many potential applications including scanning in remote communities without access to experienced sonographers and point of care focused cardiac ultrasound.
Author Disclosures: C. Luong: None. A. Abdi: None. J. Jue: None. K. Gin: None. S. Fleming: None. P. Abolmaesumi: None. T. Tsang: None.
- © 2016 by American Heart Association, Inc.