| Zugriffsnummer | 50372 |
| Dokumenttyp | Zeitschriftenartikel |
| Peer Review | mit Peer Review |
| Sprache | Englisch |
| Titel | Multilevel comparison of deep learning models for function quantification in cardiovascular magnetic resonance: On the redundancy of architectural variations |
| Autor(in); Institution |
Ammann, C.; Working Group on CMR, Experimental and Clinical Research Center, A cooperation between the Max Delbrück Center for Molecular Medicine in the Helmholtz Association and Charité — Universitätsmedizin Berlin, Berlin, GERMANY; Charité — Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, GERMANY; Max Delbrück Center for Molecular Medicine in the Helmholtz Association (MDC), Berlin, GERMANY
Hadler, T.; Working Group on CMR, Experimental and Clinical Research Center, A cooperation between the Max Delbrück Center for Molecular Medicine in the Helmholtz Association and Charité — Universitätsmedizin Berlin, Berlin, GERMANY; Charité — Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, GERMANY; Max Delbrück Center for Molecular Medicine in the Helmholtz Association (MDC), Berlin, GERMANY
Gröschel, J; Working Group on CMR, Experimental and Clinical Research Center, A cooperation between the Max Delbrück Center for Molecular Medicine in the Helmholtz Association and Charité — Universitätsmedizin Berlin, Berlin, GERMANY; Charité — Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, GERMANY; Max Delbrück Center for Molecular Medicine in the Helmholtz Association (MDC), Berlin, GERMANY
Schulz-Menger, J.; Working Group on CMR, Experimental and Clinical Research Center, A cooperation between the Max Delbrück Center for Molecular Medicine in the Helmholtz Association and Charité — Universitätsmedizin Berlin, Berlin, GERMANY; Charité — Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, GERMANY; Max Delbrück Center for Molecular Medicine in the Helmholtz Association (MDC), Berlin, GERMANY
|
| Quelle/Jahr | Frontiers in Cardiovascular Medicine: 10 (2023), 1 - 14 |
| Availability | [online only] |
| ISSN | 2297-055X (online) |
| DOI | |
| URL | |
| Verlag | Lausanne: Frontiers |
| Zusammenfassung | Cardiovascular magnetic resonance (CMR) is considered the gold standard for an accurate and reproducible assessment of cardiac anatomy and function (1, 2). Furthermore, CMR is unique in noninvasive imaging for its capabilities to characterize myocardial tissue (3) and is increasingly being included in clinical guidelines (4–6). Quantitative clinical parameters for ventricular function such as end-diastolic and end-systolic volumes, ejection fraction and left ventricular myocardial mass are predictive of patient outcome and relevant for treatment (6). Their calculation depends on exact contouring of ventricular blood volumes and myocardium. Manual segmentation is time-consuming and typically takes trained physicians up to 20 min per subject (7). In recent years convolutional neural networks (CNN) demonstrated promising results for automating semantic segmentation tasks in the medical domain (8, 9). Next to a substantial time advantage, the reproducibility of automatic image analysis eliminates the interobserver error between different readers and the intraobserver variability for the same reader at different times. Deep learning-based methods are easy to deploy to medical image segmentation tasks as they do not require geometric a-priori-knowledge or extensive feature engineering. Automated deep learning approaches match or exceed the performance of established conventional algorithms, typically measured by total segmentation overlap and mean differences in clinical parameters. Despite published overall results in the range of interobserver errors for cardiac function quantification (7, 10, 11), however, CNNs continue to make errors that compromise their acceptance for clinical application (10, 12) as generalizability and reliability remain challenging (13). Errors are not necessarily reflected in the overall results of the method, but they violate anatomical principles and are incomprehensible to human experts. Variations to the U-Net architecture [e.g., residual connections (14) or inception modules (15)] intend to improve robustness and accuracy. Yet, it remains questionable to what extent these modifications offer a substantial benefit to the segmentation accuracy given the increasing complexity and computational power requirements. The aim of this work is to provide a detailed analysis and comparison of three different CNN architectures for the quantification of ventricular function in short-axis cine images. |
| Kostenfreier Zugang | Open Access Gold |
| Rechteinformation | CC BY 4.0 ; Creative Commons Attribution 4.0 License |
| Themenbereich der Metrologie | Metrologie in der Medizin |
Zitierung
Ammann, C., Hadler, T., Gröschel, J., Kolbitsch, C., & Schulz-Menger, J. (2023). Multilevel comparison of deep learning models for function quantification in cardiovascular magnetic resonance: On the redundancy of architectural variations. Frontiers in Cardiovascular Medicine, 10, 1–14. https://doi.org/10.3389/fcvm.2023.1118499