Utility of unsupervised deep learning using a 3D variational autoencoder in detecting inner ear abnormalities on CT...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Inner ear malformations present diagnostic challenges.
• Accurate detection and localization are crucial for patient management.
• Current imaging techniques may have limitations in identifying subtle abnormalities.

Purpose of the Study:

• To evaluate the diagnostic performance of unsupervised deep learning using a 3D variational autoencoder (VAE) for detecting and localizing inner ear abnormalities.
• To compare the efficacy of unsupervised 3D-VAE with supervised deep learning methods.

Main Methods:

• Analysis of 6663 temporal bone CT images (normal and malformations).
• Unsupervised learning using 3D-VAE on a subset of images, generating difference maps.
• Supervised learning using a 3D deep residual network with 10-fold cross-validation.

Main Results:

• Unsupervised 3D-VAE achieved high diagnostic accuracy (AUC 0.99, 92.0% specificity, 99.1% sensitivity).
• The 3D-VAE method effectively highlighted abnormal regions in most malformation cases.
• Supervised learning showed high specificity (99.8%) but lower sensitivity (90.3%) and less clear localization of abnormalities.

Conclusions:

• Unsupervised deep learning with 3D-VAE offers precise detection and localization of inner ear malformations.
• The 3D-VAE approach provides a clear basis for diagnosis, unlike supervised methods in some cases.
• This AI-driven method shows significant promise for improving the diagnosis of inner ear abnormalities.