Hybrid-structure network and network comparative study for deep-learning-based speckle-modulating optical coherence...

Area of Science:

• Biomedical Imaging
• Optical Engineering
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is a noninvasive technique for 3D microstructure imaging.
• Speckle noise significantly degrades OCT image quality and resolution.
• Existing speckle reduction methods have limitations in handling diverse speckle patterns.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for speckle reduction in OCT.
• To compare the performance of various deep learning architectures for speckle pattern analysis and removal.
• To introduce a custom-built speckle-modulating OCT system and dataset for comprehensive evaluation.

Main Methods:

• Implementation of a hybrid-structure network: residual-dense-block U-Net generative adversarial network (RDBU-Net GAN).
• Creation of a custom dataset with general speckle patterns from a novel speckle-modulating OCT.
• Comparative analysis of multiple deep learning architectures for speckle extraction and removal.

Main Results:

• The proposed RDBU-Net GAN demonstrated superior performance in extracting speckle characteristics.
• RDBU-Net GAN achieved significant speckle removal, leading to improved microstructure resolution.
• The study confirmed the effectiveness of the RDBU-Net GAN over other architectures on the custom dataset.

Conclusions:

• Deep learning, particularly the RDBU-Net GAN, offers a powerful solution for OCT speckle noise reduction.
• The developed speckle-modulating OCT and dataset enable robust evaluation of deep learning models.
• This work paves the way for advanced deep-learning-based OCT systems with enhanced imaging capabilities.