Flexible Point Cloud Alignment for Middle Ear Diagnosis via Endoscopic OCT

Overview

This study presents a novel method combining ex vivo micro-CT middle ear models with in vivo endoscopic OCT data using a neural network-based non-rigid registration pipeline. The approach improves interpretation of noisy and partial OCT volumetric data by aligning it with a complete template, enabling better visualization of middle ear structures.

Background

The middle ear comprises the tympanic membrane and ossicle chain, crucial for impedance matching between air and the inner ear. Middle ear disorders such as otitis media and trauma can cause structural deformities leading to hearing loss. Conventional diagnostic tools are limited in identifying the precise origin of transmission loss. Endoscopic optical coherence tomography (OCT) offers high-resolution, depth-resolved imaging but suffers from noise and incomplete visualization of deeper structures like ossicles.

Data Highlights

The study utilizes ex vivo middle ear models reconstructed from micro-CT scans as templates and in vivo OCT scans converted into point clouds. Synthetic datasets simulating noisy and partial point clouds were generated using Blender3D to train the neural network. The registration pipeline, C2P-Net, effectively aligns complete template point clouds to partial in vivo OCT data, overcoming challenges of noise and incompleteness.

Key Findings

• A Blender3D pipeline was developed to simulate realistic synthetic shape variants and noisy partial point clouds resembling in vivo OCT data.
• C2P-Net, a two-stage non-rigid registration neural network, robustly aligns complete ex vivo templates to partial in vivo OCT point clouds.
• The neural network was trained on synthetic data with random noise and partiality, enabling generalization to patient-specific OCT scans.
• The approach improves identification of deeper middle ear structures such as the incus and stapes, which are difficult to visualize with OCT alone.
• The method leverages recent advances in point cloud feature extraction and registration, combining NgeNet and NDP architectures for effective non-rigid alignment.

Clinical Implications

This flexible point cloud alignment method enhances the diagnostic capability of endoscopic OCT by providing clearer 3D visualization of middle ear anatomy, including ossicles. Improved interpretation of OCT data may aid in more accurate diagnosis of middle ear pathologies and better understanding of conductive hearing loss origins. The approach could complement existing diagnostic tools by localizing structural abnormalities non-invasively.

Conclusion

The integration of ex vivo micro-CT templates with in vivo OCT data via a neural network-based non-rigid registration pipeline represents a significant advancement in middle ear imaging. This method addresses limitations of OCT alone and holds promise for improved clinical assessment of middle ear conditions.

References

1. Middle Ear Function and Pathology Overview
2. Endoscopic OCT for Middle Ear Imaging
3. Point Cloud Registration Techniques and Neural Networks