Cross-attention learning enables real-time nonuniform rotational distortion correction in OCT

Area of Science:

• Medical Imaging
• Optical Coherence Tomography
• Artificial Intelligence

Background:

• Nonuniform rotational distortion (NURD) correction is critical for accurate endoscopic optical coherence tomography (OCT) imaging.
• Existing NURD correction techniques are computationally intensive, limiting real-time applications and temporal resolution.

Purpose of the Study:

• To introduce a novel, efficient cross-attention learning method for NURD correction in endoscopic OCT.
• To accelerate NURD correction while maintaining or improving image quality.

Main Methods:

• Developed an end-to-end stacked cross-attention network inspired by natural language processing and computer vision.
• Incorporated three types of optimization constraints to enhance correction performance.
• Evaluated the method on public endoscopic OCT datasets and a 3D stent reconstruction system.

Main Results:

• Achieved a ~3x speedup, reaching real-time performance (26±3 fps).
• Demonstrated superior NURD correction compared to traditional feature-based and CNN-based methods.
• Validated effectiveness on 3D stent reconstruction using a custom endoscopic OCT system.

Conclusions:

• The proposed cross-attention learning method offers a significant advancement in NURD correction for endoscopic OCT.
• This technique enables real-time processing and enhances the accuracy of OCT imaging and its extensions.