Deep Learning System for Noise-Free Optical Coherence Tomography (OCT) Imaging without Clean Training Data
This is an innovative deep learning strategy called Sub2Full (S2F) that significantly reduces speckle noise in Optical Coherence Tomography (OCT) images. Unlike conventional methods, this system achieves superior image quality without the difficult-to-obtain "clean" training data, allowing for the visualization of fine sublaminar anatomical structures.
Description
The S2F system is a deep learning-based denoising scheme designed to overcome a major limitation of supervised methods: the requirement for noise-free images for network training. The innovative core of S2F is its "split spectrum to full spectrum" training process, which utilizes only noisy data. The method works by acquiring two repeated OCT scans (B-scans) of the same object. The key feature is the unique input/output mapping: a low-resolution (LR) noisy image is generated by splitting the spectrum of the first scan (R1) to serve as the network's input, while the high-resolution (HR), full-spectrum image of the second scan (R2) is set as the noisy target. This novel approach, which creates a one-to-many (1:N) mapping of the LR input to multiple potential HR outcomes, forces the deep learning network's optimization to converge toward the true clean image, yielding superior despeckling results compared to existing techniques.Applications
- Ophthalmology Diagnostics: Enhanced imaging for early and accurate diagnosis of retinal diseases requiring visualization of fine sub-retinal structures.- OCT Device Manufacturing: Integration into the image processing pipeline of new or existing OCT devices to boost image output quality.
- Intravascular Imaging: Improved clarity of catheter-based OCT to better identify and characterize arterial plaques and vessel walls.
- Dermatology and Sub-dermal Imaging: Non-invasive, high-resolution cross-sectional imaging for skin conditions.
- Preclinical Research: Higher quality imaging for translational studies and drug development in animal models.
Advantages
- Eliminates the need for clean data by successfully training the deep learning model using only two sets of noisy OCT scans, simplifying data acquisition.- Provides superior image quality and more effective speckle noise reduction than current state-of-the-art methods like Noise2Noise (N2N) and Noise2Void (N2V).
- Enables the visualization of previously obscured fine anatomical structures, such as sublaminar layers in the outer retina, which are critical for high-resolution diagnostics.
- Has the potential to be applied to all types of Optical Coherence Tomography (OCT) systems, including high-resolution modalities like visible light OCT.