Synthetic polarization-sensitive optical coherence tomography using contrastive unpaired translation

Area of Science:

• Biomedical optics
• Medical imaging
• Tissue birefringence analysis

Background:

• Polarization-sensitive optical coherence tomography (PS-OCT) is crucial for analyzing biological tissue birefringence.
• Generating high-fidelity PS-OCT images traditionally requires extensive labeled datasets.
• Existing methods struggle to efficiently replicate periodic patterns in PS-OCT images.

Purpose of the Study:

• To develop an efficient method for generating synthetic PS-OCT images from single OCT images using Contrastive Unpaired Translation (CUT).
• To address the limitations of data-intensive, pixel-wise correlation methods in capturing birefringence patterns.
• To evaluate the performance of CUT against other generative models for PS-OCT image synthesis.

Main Methods:

• Employed Contrastive Unpaired Translation (CUT), a generative model leveraging patch-wise correlations on unpaired data.
• Compared CUT with Pix2pix and CycleGAN on in vivo mouse tendon healing data over six weeks.
• Utilized a ResNet-152 model for tissue damage classification based on generated PS-OCT images.

Main Results:

• CUT successfully generated high-fidelity synthetic PS-OCT images, closely matching original images.
• CUT demonstrated superior performance (p < 0.0001) compared to Pix2pix and CycleGAN, confirmed by statistical tests.
• The ResNet-152 model achieved up to 90.13% accuracy in tissue damage assessment using synthetic PS-OCT images.

Conclusions:

• Contrastive Unpaired Translation (CUT) is a highly effective and efficient method for synthesizing PS-OCT images.
• CUT accurately captures underlying tissue structural features responsible for birefringence.
• The proposed approach offers significant improvements in image fidelity and efficiency for PS-OCT analysis.