Deep learning generalization study on optical coherence tomography image denoising.

Area of Science:

• Medical Imaging
• Optical Coherence Tomography (OCT)
• Deep Learning

Background:

• Image quality in OCT is significantly impacted by noise.
• Current deep learning denoising methods struggle with generalization to unseen noise levels.

Purpose of the Study:

• To enhance the adaptability of deep learning denoising models to varying noise conditions in OCT.
• To improve the generalization capability of OCT denoising models.

Main Methods:

• Developed a mixed-training strategy using a multi-noise level dataset for SS-OCT.
• Constructed datasets with simulated noise using optical attenuators (4 dB, 6 dB, 10 dB).
• Applied the strategy to various denoising networks (ResNet, U-Net, DnCNN, AG-DCN) under supervised and unsupervised learning.

Main Results:

• Models trained with the mixed-training strategy demonstrated robust performance across different noise levels, including unseen 4 dB noise.
• The supervised U-Net model achieved a PSNR of 29.233 dB and SSIM of 0.807 on 4 dB data, comparable to dedicated models.
• Mixed-trained networks effectively suppressed noise artifacts, validating their performance under mismatched noise conditions.

Conclusions:

• Multi-noise level datasets are valuable for improving deep learning model generalization.
• The proposed mixed-training strategy enhances adaptability and supports reliable OCT analysis in practical applications.