Optimizing 3D FLAIR imaging of endolymphatic hydrops: Benefits and limitations of deep learning reconstruction.
Authors
Affiliations (7)
Affiliations (7)
- Guilloz Imaging Department, Central Hospital, University Hospital Center of Nancy, 54000 Nancy, France.
- Université de Lorraine, INSERM, IADI, 54000 Nancy, France; Université de Lorraine, CIC, Innovation Technologique, University Hospital Center of Nancy, 54000 Nancy, France.
- ENT Surgery Department, Central Hospital, University Hospital Center of Nancy, 54000, Nancy, France; Université de Lorraine, DevAh, 54500, Vandoeuvre-Les-Nancy, France.
- Siemens Healthineers AG, 91301, Forchheim, Germany.
- Guilloz Imaging Department, Central Hospital, University Hospital Center of Nancy, 54000 Nancy, France; Université de Lorraine, INSERM, IADI, 54000 Nancy, France; Université de Lorraine, CIC, Innovation Technologique, University Hospital Center of Nancy, 54000 Nancy, France.
- Clinique de l'Oreille, 75001, Paris, France.
- Guilloz Imaging Department, Central Hospital, University Hospital Center of Nancy, 54000 Nancy, France; Université de Lorraine, INSERM, IADI, 54000 Nancy, France; Université de Lorraine, CIC, Innovation Technologique, University Hospital Center of Nancy, 54000 Nancy, France. Electronic address: [email protected].
Abstract
The purpose of this study was to evaluate the influence of deep learning reconstruction (DLR) on image quality and diagnostic performance of three-dimensional (3D) fluid-attenuated inversion recovery (FLAIR) magnetic resonance imaging (MRI) of the inner ear. Fifty-three patients (32 women; mean age, 57.4 ± 16.1 [standard deviation (SD)] years) were prospectively included. All patients underwent 3D FLAIR MRI of the inner ear. Four MRI sequences were analyzed, including conventional FLAIR (acquisition time, 7 min 14 s), high-resolution FLAIR (6 min 54 s), DLR-FLAIR (6 min 54 s), and accelerated DLR-FLAIR (5 min 34 s). Signal-to-noise ratio (SNR) was measured for objective analysis. Two readers independently evaluated diagnostic performance using conventional 3D FLAIR images as the standard of reference, and image quality (i.e., motion artifacts, noise, edge sharpness, global quality). Interobserver agreement was assessed using kappa statistics and Bland Altman analysis. DLR significantly increased SNR compared to non-DLR sequences (P < 0.001), with the highest SNR values observed for DLR-FLAIR images (91.2 ± 29.2 [SD]). Subjective analysis showed significant noise reduction and improved global image quality with DLR-FLAIR images (P < 0.001). Sensitivities for the diagnosis of inner-ear findings of the different sequences ranged between 62.5% and 100% and specificities between 92.1% and 100%, with no significant differences between sequences. Interobserver agreement was substantial to almost perfect for most findings but decreased for subtle structures and with the accelerated DLR-FLAIR sequence. Bland Altman analysis showed no systematic bias but increased variability with DLR. DLR significantly improves image quality in 3D FLAIR imaging of the inner ear, without compromising diagnostic performance. However, it may influence interobserver variability, particularly for subtle findings and accelerated acquisitions, which may modify diagnostic behavior, highlighting the importance of adapted interpretation strategies.