Texture preservation and speckle reduction in poor optical coherence tomography using the convolutional neural...

Area of Science:

• Medical Imaging
• Image Processing
• Deep Learning

Background:

• Optical coherence tomography (OCT) images often suffer from poor quality due to speckle noise, edge blur, and texture loss.
• These artifacts are particularly problematic in background regions near edges, hindering accurate analysis.

Purpose of the Study:

• To propose a novel de-speckling method for enhancing the quality of poor-quality OCT images.
• To develop a deep convolutional neural network (CNN) model capable of effectively removing speckle noise while preserving image details.

Main Methods:

• A deep nonlinear CNN mapping model, named OCTNet, was developed in a serial architecture.
• A pertinent dataset was constructed by combining three existing methods to train the OCTNet model.
• The proposed method accurately extracts speckle noise from original OCT images.

Main Results:

• OCTNet effectively reduces speckle noise, protects structural information, and preserves edge features simultaneously, even near edges.
• Quantitative and qualitative evaluations showed superior performance compared to adaptive complex diffusion, curvelet shrinkage, and shearlet-based total variation methods.
• The method demonstrated excellent generalization, adaptiveness, robustness, and batch performance.

Conclusions:

• The proposed OCTNet method significantly enhances OCT image quality by effectively reducing speckle noise.
• OCTNet is suitable for rapid processing of diverse OCT images in real-time situations without parameter fine-tuning.
• This deep learning approach offers a robust solution for improving OCT image analysis in medical diagnostics.