Real-time noise reduction based on ground truth free deep learning for optical coherence tomography

Area of Science:

• Biomedical Imaging
• Medical Image Processing
• Deep Learning

Background:

• Optical coherence tomography (OCT) is a crucial non-invasive 3D imaging technique in biomedical research and clinical settings.
• Image noise is an inherent challenge in OCT, hindering accurate post-processing and diagnosis.
• Existing noise reduction methods, like frame averaging and deep learning with ground truth, have limitations such as long acquisition times or complex data preparation.

Purpose of the Study:

• To develop and evaluate a deep learning-based noise reduction method for OCT images that does not require clean reference images for training.
• To compare the performance of three different neural network architectures (Unet, SRResNet, AC-SRResNet) for OCT noise reduction.
• To assess the effectiveness and real-time applicability of the trained models in an accelerated OCT imaging system.

Main Methods:

• Three deep learning network structures—Unet, super-resolution residual network (SRResNet), and asymmetric convolution-SRResNet (AC-SRResNet)—were trained for noise reduction.
• Models were evaluated using metrics including signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and edge preservation index (EPI).
• The trained models were integrated into a GPU-accelerated OCT imaging system for real-time noise reduction performance evaluation.

Main Results:

• All three trained models (Unet, SRResNet, AC-SRResNet) demonstrated significant SNR improvements, with L2-loss trained models achieving up to 24.88 dB.
• AC-SRResNet and SRResNet showed superior denoising compared to Unet, albeit with longer computation times.
• AC-SRResNet exhibited better edge preservation than SRResNet, while Unet's performance was comparable.
• Real-time noise reduction was achieved at high frame rates (up to 64 fps for Unet, 19 fps for SRResNet, 17 fps for AC-SRResNet) for 512x512 pixel OCT images.

Conclusions:

• A novel deep learning approach effectively reduces noise in OCT images without requiring clean ground truth data.
• The evaluated deep learning models offer a viable solution for enhancing OCT image quality, with AC-SRResNet providing a good balance of denoising and edge preservation.
• The integration of these models into an accelerated OCT system enables real-time noise reduction, advancing the utility of OCT in clinical and research applications.