| Datensatz noch nicht abgeschlossen | |
| Zugriffsnummer | 56591 |
| Dokumenttyp | Konferenzartikel |
| Sprache | Englisch |
| Titel | Model-based deep learning MRI reconstruction: adversarial robustness vs. network capacity |
| Autor(in); Institution |
Bernhardt, Bill A.; 8.1, Biomedizinische Magnetresonanz, PTB-Berlin
|
| Quelle/Jahr | Proceedings of the International Society for Magnetic Resonance in Medicine:(2026) |
| Availability | [online only] |
| ISSN | 1065-9889 |
| DOI | |
| Verlag | ISMRM |
| Konferenzangaben | Annual Meeting of the ISMRM, Capetown, 9-14, Mai, 2026, South Africa |
| Freie Schlagworte | Image Reconstruction ; Deep Learning ; Adversarial Attacks ; Adversarial Robustness ; Adversarial Stability |
| Zusammenfassung | Motivation: Deep-learning-based regularization forms the modern state of the art for MR image reconstruction, leveraging the image-processing capabilities of neural networks and incorporating the physical model of MRI acquistions. Typically, employing a higher number of trainableparameters (NTP) in the network yields empirically better performance with respect to the achievable image metrics. Goals:To examine a potentially inherent relationship between network size and the stability of reconstruction networks with respect to adversarialperturbations. Approach:We performed adversarial attacks on networks with varying NTP. Results:Larger networks tend to exhibit greater sensitivity to adversarial perturbations. This effect is exacerbated by suboptimal network training. |
