Probabilistic volumetric speckle suppression in OCT using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Speckle noise in Optical Coherence Tomography (OCT) degrades image quality and hinders accurate analysis.
• Developing effective speckle reduction techniques is crucial for advancing OCT applications in various biomedical fields.

Purpose of the Study:

• To introduce a novel deep learning framework for volumetric speckle reduction in OCT data.
• To leverage the volumetric nature of OCT data for enhanced speckle suppression and resolution preservation.

Main Methods:

• A conditional generative adversarial network (cGAN) was designed to process partial OCT volumes.
• Volumetric non-local means (TNode) was used to generate high-quality training data for the cGAN.
• The framework was trained on minimal data (three OCT volumes) and validated across different OCT systems and tissue types.

Main Results:

• The cGAN achieved significant speckle reduction and preserved tissue structures across all three dimensions.
• The proposed method demonstrated a speed improvement of two orders of magnitude compared to TNode.
• The framework showed effective performance on unseen tissue types, highlighting its generalizability.

Conclusions:

• The developed deep learning framework offers a fast and effective solution for volumetric speckle reduction in OCT.
• The open-source, all-software implementation facilitates widespread adoption and adaptation for diverse OCT systems.
• This work addresses the challenge of training data generation, enabling high-quality despeckling without extensive ground truth datasets.