[Paper Review] A resnet-based universal method for speckle reduction in optical coherence tomography images

Ning Cai, Fei Shi+2 more|arXiv (Cornell University)|Mar 22, 2019

Optical Coherence Tomography Applications11 references4 citations

TL;DR

This paper proposes a ResNet-based deep learning method for universal speckle reduction in optical coherence tomography (OCT) images, leveraging a novel architecture with Convolution-BN-ReLU, Branch, and Residual modules to automatically learn noise characteristics without manual parameter tuning. The method achieves over 22 dB improvement in signal-to-noise ratio with minimal structural distortion and generalizes across diverse OCT image types without retraining.

ABSTRACT

In this work we propose a ResNet-based universal method for speckle reduction in optical coherence tomography (OCT) images. The proposed model contains 3 main modules: Convolution-BN-ReLU, Branch and Residual module. Unlike traditional algorithms, the model can learn from training data instead of selecting parameters manually such as noise level. Application of this proposed method to the OCT images shows a more than 22 dB signal-to-noise ratio improvement in speckle noise reduction with minimal structure blurring. The proposed method provides strong generalization ability and can process noisy other types of OCT images without retraining. It outperforms other filtering methods in suppressing speckle noises and revealing subtle features.

Motivation & Objective

• To address the challenge of speckle noise degradation in OCT imaging, which hinders diagnostic accuracy and feature visibility.
• To develop a deep learning-based method that automatically learns noise patterns without manual parameter tuning, such as noise level estimation.
• To create a universal model capable of processing various OCT image types without retraining, ensuring broad clinical applicability.
• To outperform traditional filtering methods in suppressing speckle noise while preserving fine anatomical structures.

Proposed method

• The proposed model integrates three core modules: Convolution-Batch Normalization-ReLU (Conv-BN-ReLU), a Branch module for multi-scale feature extraction, and a Residual module for learning residual noise patterns.
• The architecture is based on ResNet, enabling deep network training with skip connections that mitigate vanishing gradient problems.
• The model is trained end-to-end on OCT image patches, learning to map noisy inputs to clean outputs through supervised learning.
• The Branch module enhances multi-scale representation learning, improving sensitivity to subtle structural details in noisy OCT data.
• The Residual learning strategy allows the network to focus on learning the difference between noisy and clean images, improving convergence and performance.
• The method does not require explicit noise level estimation during inference, enabling universal application across different OCT acquisition settings.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on OCT data achieve superior speckle reduction compared to traditional filtering techniques?
• RQ2 To what extent can a single deep learning model generalize across diverse OCT image types without retraining?
• RQ3 Does the proposed ResNet-based architecture preserve fine anatomical structures while reducing speckle noise?
• RQ4 How does the model perform in terms of signal-to-noise ratio (SNR) improvement without manual parameter tuning?

Key findings

• The proposed method achieves a signal-to-noise ratio (SNR) improvement of over 22 dB in speckle noise reduction, significantly outperforming conventional filtering methods.
• The model preserves fine anatomical structures with minimal blurring, maintaining diagnostic image quality.
• The method generalizes effectively across different types of OCT images without requiring retraining, demonstrating strong universal applicability.
• The model outperforms existing filtering techniques in both noise suppression and feature recovery, particularly for subtle structural details.
• The absence of manual parameter tuning, such as noise level estimation, enables plug-and-play deployment in diverse clinical settings.