Live 4D-OCT denoising with self-supervised deep learning.

Area of Science:

• Ophthalmic imaging
• Medical technology
• Artificial intelligence in medicine

Background:

• Live volumetric optical coherence tomography (4D-OCT) offers high-resolution 3D visualization for ophthalmic surgery.
• High imaging speeds in 4D-OCT increase noise, limiting real-time applications.
• Current noise reduction methods face limitations due to high data rates and latency requirements.

Purpose of the Study:

• To propose and evaluate a low-complexity neural network for real-time denoising in 4D-OCT.
• To integrate the neural network into the image reconstruction pipeline of a 4D-OCT prototype.
• To improve the visual quality of volumetric renderings for enhanced intra-surgical guidance.

Main Methods:

• Developed a blind-spot neural network using a self-supervised learning approach on unpaired OCT images.
• Integrated an optimized U-Net architecture into a 1.2 MHz A-scan rate microscope-integrated 4D-OCT prototype.
• Compared the neural network's denoising performance against non-local filtering and Gaussian filtering algorithms.

Main Results:

• The neural network introduced only a few milliseconds of additional latency.
• Achieved superior denoising performance compared to the basic setup and non-local filtering.
• Preserved anatomical structure details (layers and edges) better than Gaussian filtering with comparable processing time.
• Demonstrated improved visual appearance of volumetric renderings through real-time denoising.

Conclusions:

• A low-complexity, self-supervised neural network can effectively denoise 4D-OCT images in real-time.
• This approach enhances the visual quality of intra-surgical renderings, crucial for clinical acceptance.
• The developed method represents a significant step towards the translation of 4D-OCT as a reliable surgical guidance tool.