A Deep Learning Framework for Echocardiographic View Classification Using Spatial, Temporal, and Radon Features Publisher



Mohammadi M ; Talebpour A ; Hosseinsabet A
Source: Healthcare Analytics Published:2026

Abstract

This study introduces a novel analytics-driven approach for classifying transthoracic echocardiographic (TTE) views by integrating hybrid deep learning, transfer learning, and a three-stream network architecture. The proposed method combines spatial features, temporal dynamics, and radon-transformed representations to enhance model performance and distinguish visually similar echocardiographic views. Trained and evaluated on the EchoIR database, which contains 5901 videos from 757 patients across 24 view classes, including color Doppler views, this model achieved classification accuracies of 94.61 ± 0.89% and 95.53 ± 0.67% on two benchmark evaluation protocols. By embedding both temporal and Radon-based features, the approach demonstrates significant improvements in predictive accuracy and computational efficiency over existing methods. In the clinical context, accurate view classification is essential for reliable labeling, dataset construction, and the development of automated diagnostic tools. By providing accurate predictions across a large set of echocardiograms, the proposed method facilitates scalable and reliable cardiovascular diagnostic workflows. © 2026 Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/4.0/