Deep learning for discriminating cochlear malformations on temporal bone CT.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Diagnosing cochlear malformations on temporal bone CT scans presents challenges due to subtle imaging findings.
• Accurate diagnosis is crucial for appropriate patient management and surgical planning.

Purpose of the Study:

• To evaluate the effectiveness of deep learning (DL) algorithms in identifying cochlear malformations using temporal bone CT images.
• To compare the diagnostic performance of DL models against human expert (otologist) interpretation.

Main Methods:

• A dataset of 373 temporal bone CTs was utilized, with cochleae automatically segmented using the Swin UNETR network.
• Three classification networks (ResNet50, EfficientNet-B0, DenseNet121) were trained and tested on 590 cochlear sides.
• Performance was evaluated using the Area Under the Curve (AUC) metric and compared to otologist diagnoses.

Main Results:

• The Swin UNETR model achieved an average Dice coefficient of 0.93 for cochlear segmentation.
• DenseNet121 and ResNet50 achieved the highest AUCs (0.93), outperforming EfficientNet-B0 (0.89).
• DL models, particularly ResNet50 and DenseNet121, demonstrated superior diagnostic accuracy compared to otologists with 1-5 years of experience.

Conclusions:

• Deep learning analysis offers a valuable supplementary approach for diagnosing cochlear malformations on temporal bone CT.
• AI-powered tools have the potential to enhance diagnostic accuracy and efficiency in otologic imaging interpretation.