Multiscale denoising generative adversarial network for speckle reduction in optical coherence tomography images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Optical coherence tomography (OCT) is a high-resolution, noninvasive imaging technique for visualizing tissue microstructures.
• Speckle noise is an inherent artifact in OCT images, degrading image quality and hindering accurate disease diagnosis.
• Effective speckle reduction methods are crucial for improving the clinical utility of OCT.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for reducing speckle noise in OCT images.
• To enhance the quality and diagnostic potential of OCT imaging through advanced denoising techniques.

Main Methods:

• A multiscale denoising generative adversarial network (MDGAN) was proposed, incorporating a cascade multiscale module and a spatial attention mechanism.
• A deep back-projection layer was utilized for enhanced feature learning within the MDGAN architecture.
• The MDGAN model was trained and validated on two distinct OCT image datasets.

Main Results:

• MDGAN demonstrated significant improvements in peak signal-to-noise ratio and signal-to-noise ratio, achieving up to a 3 dB enhancement.
• While structural similarity index measurement and contrast-to-noise ratio were marginally lower than the best existing methods (1.4% and 1.3% respectively), overall performance was superior.
• Experimental results confirmed the effectiveness of MDGAN in speckle reduction across various scenarios.

Conclusions:

• The proposed MDGAN is an effective and robust method for speckle reduction in OCT images.
• MDGAN outperforms existing state-of-the-art denoising techniques, offering substantial improvements in image quality.
• This advancement has the potential to significantly improve OCT imaging-based disease diagnosis by mitigating speckle artifacts.