Synchrotron-Based Deep Learning Network of the Inner Ear: Development and Expert Validation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anatomy

Background:

• Accurate segmentation of the inner ear is crucial for diagnosing and treating various otological conditions.
• Manual segmentation of the inner ear on computed tomography (CT) scans is time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop and validate a deep learning (DL) network for automated inner ear segmentation in clinical CT scans.
• To compare the DL network's performance against domain experts and consensus segmentation methods.

Main Methods:

• A DL segmentation network was trained on 4,784 paired synchrotron-radiation phase contrast imaging (SR-PCI) and clinical CT datasets.
• The network was developed using 100 cadaveric specimens with diverse CT acquisition protocols and resolutions.
• External validation was performed against seven domain experts and STAPLE consensus segmentation on an unseen dataset.

Main Results:

• The DL network achieved a Dice similarity coefficient of 0.922 and Hausdorff distances of 0.329 mm (max) and 0.006 mm (avg) compared to SR-PCI ground truth.
• The automated segmentation significantly outperformed individual experts, average expert performance, and STAPLE consensus.
• The network demonstrated high accuracy on cone-beam CT and helical CT with resolutions down to 625 μm.

Conclusions:

• This study presents the first automated inner ear segmentation algorithm that surpasses expert performance.
• The developed DL network establishes a new potential clinical gold standard for inner ear segmentation.
• Automated segmentation offers a reliable and efficient alternative to manual delineation in clinical practice.