C2P-Net

C2P-Net performs two-stage non-rigid point cloud registration to align ex vivo middle ear models with in vivo endoscopic optical coherence tomography (OCT) point clouds and thereby improve interpretability of OCT for middle ear diagnostics.

Key Features:

• Two-staged non-rigid registration: Implements a two-stage pipeline for non-rigid registration of point clouds derived from ex vivo middle ear models and in vivo OCT.
• Point cloud inputs: Operates on noisy, partial, and volumetric point clouds representing middle ear structures from endoscopic OCT and ex vivo models.
• Blender3D-based training data generation: Uses Blender3D to simulate various middle ear shapes and produce noisy and partial point clouds for training data.
• Morphological and volumetric integration: Merges morphological information from ex vivo models with volumetric OCT data to enhance OCT image interpretability.
• Robustness to occlusion and incompleteness: Explicitly handles occlusions, realistic noise, and incompleteness in point cloud data.
• Evaluation on synthetic and real data: Validated on synthetic datasets and actual endoscopic OCT data demonstrating generalizability to unseen middle ear point clouds.

Scientific Applications:

• Middle ear infection diagnostics: Improves interpretation of endoscopic OCT for assessment of middle ear infections, including cases prevalent in pediatric populations.
• Clinical otology imaging: Integrates ex vivo morphological models with in vivo OCT to support otologists in identifying middle ear pathologies from volumetric OCT point clouds.
• Training data augmentation for machine learning: Provides simulated noisy and partial point clouds to mitigate limited labeled training data for ML applications on middle ear OCT.

Methodology:

Two-staged non-rigid registration of point clouds; synthetic training data generation using Blender3D to simulate noisy and partial middle ear point clouds; evaluation on synthetic datasets and actual endoscopic OCT data.