Application of UNETR for automatic cochlear segmentation in temporal bone CTs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anatomy

Background:

• Accurate segmentation of the cochlea is crucial for diagnosing inner ear pathologies.
• Manual segmentation is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To evaluate the feasibility of a UNETR-based deep learning model for automated cochlear segmentation.
• To compare the UNETR model's performance against a 3D U-Net model.

Main Methods:

• Utilized temporal bone CT scans from 77 patients.
• Applied both 3D U-Net and UNETR models for automatic cochlear segmentation.
• Tested models on normal, GE 256 CT, SE-DS CT, and cochlear deformity datasets.

Main Results:

• The UNETR model achieved a Dice coefficient of 0.92 on the normal cochlear test set, outperforming the 3D U-Net model.
• Dice coefficients for UNETR on GE 256 CT, SE-DS CT, and Cochlear Deformity CT datasets were 0.91, 0.93, and 0.93, respectively.
• Optimal performance was observed with batch_size=1.

Conclusions:

• The UNETR model enables fully automatic and accurate cochlear segmentation from temporal bone CT images.
• This deep learning approach is feasible and achieves accuracy comparable to manual segmentation.
• The method holds promise for clinical applications requiring precise cochlear analysis.