Classification, registration and segmentation of ear canal impressions using convolutional neural networks

Area of Science:

• Biomedical Engineering
• Computer Science
• Medical Imaging

Background:

• Bespoke hearing aid fitting traditionally requires manual 3D editing of silicone ear canal molds.
• Current methods for creating digital ear impressions are labor-intensive and time-consuming.
• Standardization of ear canal data is needed for morphological studies.

Purpose of the Study:

• To develop automated computational methods for pre-processing 3D ear canal impressions.
• To create tools that assist in hearing aid design, manufacturing, and fitting.
• To normalize anatomical data for the study of outer ear canal morphology.

Main Methods:

• Classification of ear canal impression handedness (left/right) using convolutional neural networks.
• Orientation of 3D geometries to a consistent coordinate system.
• Segmentation techniques to remove artifacts from silicone ear molds.
• Evaluation using a dataset of 3000 ear canal impressions.

Main Results:

• Convolutional neural networks achieved 95.8% adjusted accuracy in classification.
• Registration accuracy was within 20° angular error (92.3%).
• Segmentation performance yielded a 93.4% intersection over union.

Conclusions:

• Automated pre-processing of ear canal impressions is feasible and effective.
• Computational methods, particularly deep learning, can significantly reduce manual effort in hearing aid workflows.
• The developed techniques offer potential for improved hearing aid customization and anatomical research.