Deep Learning Reconstruction Specialized for Inner Ear: Improving Image Quality and Anatomical Structure Visualization as Compared with Conventional Hybrid-Type Iterative Reconstruction on High-Definition CT.

Background/

Objectives

To directly compare the capabilities of hybrid-type iterative reconstruction (IR) with the newly developed deep learning reconstruction (DLR) for the inner ear on high-definition CT (HDCT) obtained using the super-high-resolution (SHR) mode for external, middle and inner ear evaluations and diagnosis in patients with and without otologic diseases.

Methods

Included in this study were 140 patients who had undergone HDCT, consisting of 32 otologic disease patients and 108 non-otologic disease patients, and 280 inner and middle ears and temporal bones were evaluated on a per ear analysis. Signal-to-noise ratios (SNRs) of the temporal bone surrounding the aural vestibule of the ear and in the vestibule as well as the cerebellar hemisphere, overall image and detailed evaluation of the visibility of anatomical landmarks in the middle and inner ear and temporal bone obtained with the two methods were assessed and statistically compared using the paired t-test or Wilcoxon's signed-rank test. Then, receiver operating characteristic (ROC) analysis was performed to compare diagnostic performance between two reconstruction methods.

Results

Each SNR of DLR was significantly higher than that of hybrid-type IR (p < 0.05). Overall image quality and detailed visualization of each anatomical structure obtained with DLR were significantly better than those obtained with hybrid-type IR (p < 0.05). The area under the curve of DLR had no significant difference with hybrid-type IR (p = 0.18).

Conclusions

DLR has superior potential to hybrid-type IR for better image quality and visualization of anatomical landmarks in middle and inner ears and temporal bones on HDCT, although diagnostic performance was not affected in clinical practice.