Deep Denoising of Volumetric OCT Images for In Vivo Motion Detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Optical Coherence Tomography (OCT) offers high resolution but is degraded by speckle noise.
• Speckle noise in OCT images can cause outliers, impacting image interpretation and subsequent analyses.
• Existing denoising methods may not be suitable for real-time OCT applications.

Purpose of the Study:

• To develop and evaluate an unsupervised 3D convolutional Autoencoder (AE) for denoising OCT images.
• To assess the impact of AE-based denoising on motion detection accuracy.
• To determine the processing speed and clinical applicability of the proposed AE model.

Main Methods:

• An unsupervised 3D convolutional Autoencoder (AE) was designed for OCT image denoising.
• The AE model was systematically evaluated on both in vivo and post mortem OCT datasets.
• Motion detection performance was analyzed as a subsequent task following AE denoising.

Main Results:

• AE-based denoising effectively reduced speckle noise and outliers in OCT images.
• Denoising significantly improved the accuracy of motion detection, particularly on in vivo data.
• The proposed AE model processes OCT volumes rapidly (0.5 ms), enabling real-time applications.

Conclusions:

• Unsupervised deep learning AE provides a fast and effective method for denoising OCT images.
• AE-based denoising is crucial for mitigating artifacts and enhancing motion detection in OCT.
• The developed AE model offers a clinically relevant solution for improving 3D OCT image quality and data processing.