A deep learning approach for pose estimation from volumetric OCT data.

Area of Science:

• Medical Imaging
• Computer Vision
• Robotics

Background:

• Accurate instrument pose tracking is vital for image-guided surgery, especially in microscopic scenarios.
• Optical Coherence Tomography (OCT) offers high resolution for pose estimation but presents processing challenges like speckle noise and artifacts.
• Existing methods often rely on 2D image data, potentially limiting accuracy in 3D surgical environments.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for 6D pose estimation directly from OCT volume data.
• To investigate the efficacy of 3D Convolutional Neural Networks (CNNs) for micrometer-level pose tracking in surgical applications.
• To compare the performance of 3D CNNs against 2D approaches using depth information for pose estimation.

Main Methods:

• A new 3D CNN architecture, Inception3D, was designed for direct 6D pose prediction from OCT volumes.
• A hexapod robot was utilized for automated acquisition of labeled training data.
• The framework was trained using multi-output regression on volumetric OCT data.

Main Results:

• The 3D CNN approach demonstrated superior accuracy compared to methods using 2D representations with depth information.
• Quantitative and qualitative results confirmed that 3D CNNs effectively leverage the depth structure of marker objects.
• The proposed deep learning method achieved mean average errors of 14.89 ± 9.3 µm for position and 0.096 ± 0.072° for orientation.

Conclusions:

• Deep learning-based pose estimation using 3D OCT volumes offers high accuracy for instrument tracking in image-guided surgery.
• Exploiting volumetric information with 3D CNNs is more effective than using 2D data for precise pose estimation.
• The Inception3D architecture provides an efficient and high-performing solution for 6D pose estimation from OCT data.