Deep learning models for automated view and orientation classification and cardiac cycle phase detection in echocardiography

Szijártó, Ádám and Szeier, Thomas Akos and Kitano, Tetsuji and Nabeshima, Yosuke and Lakatos, Bálint and Fábián, Alexandra and Ladányi, Zsuzsanna and Szávai, Luca and Tolvaj, Máté and Ferencz, Andrea and Turschl, Tímea Katalin and Magyar, Bálint and Bagyura, Zsolt and Asch, Federico M. and Addetia, Karima and Takeuchi, Masaaki and Merkely, Béla Péter and Kovács, Attila and Tokodi, Márton (2026) Deep learning models for automated view and orientation classification and cardiac cycle phase detection in echocardiography. IMAGING. ISSN 2732-0960 (In Press)

Preview

Text 1647-article-10.1556-1647.2026.00404.pdf - Published Version Available under License Creative Commons Attribution. Download (1MB) | Preview

Abstract

Background Deep learning (DL) is increasingly used for the automated analysis of echocardiographic videos. These end-to-end pipelines often incorporate models targeting common preprocessing tasks to ensure the selection of videos and frames suitable for their primary analytic objective. However, publicly available models supporting these core steps remain limited. Objectives We aimed to develop three DL models for distinct preprocessing tasks: Model-V for view classification, Model-O for orientation classification, and Model-CC for identifying end-diastolic (ED) and end-systolic (ES) frames. Methods Model-V was trained and internally validated on 13,864 videos and externally evaluated on 5,650 videos. Model-O and Model-CC were trained and internally validated on 5,513 and 3,108 apical 4-chamber view (A4C) videos, respectively, and both were externally tested on the same set of 1,200 A4C videos. Results Model-V discriminated A4C from non-A4C with balanced accuracies of 0.991 (0.987–0.994) and 0.895 (0.884–0.907), and Model-O distinguished Stanford from Mayo orientation with balanced accuracies of 0.995 (0.990–1.000) and 0.960 (0.948–0.971) during internal and external testing, respectively. Model-CC identified ED frames with mean absolute errors (MAEs) of 4.041 (3.878–4.210) and 3.117 (3.044–3.193) frames and ES frames with MAEs of 2.799 (2.708–2.894) and 4.827 (4.738–4.917) frames in the internal and external test sets, while failing to identify only 7.3 and 4.2% of ED frames and 1.2 and 3.6% of ES frames, respectively. Conclusions The proposed models demonstrated good performance across key preprocessing tasks; thus, they constitute readily integrable components for end-to-end DL pipelines in automated echocardiographic analysis.

Actions (login required)

View Item