Synthetic polarization-sensitive optical coherence tomography by deep learning

The study addresses the challenge that conventional polarization-sensitive OCT (PS-OCT) requires additional detectors and polarization optics, increasing cost and complexity and complicating handheld/probe-based intraoperative use. PS-OCT provides valuable birefringence-based contrasts (e.g., phase retardation and DOPU) that aid in differentiating tissue types such as cancer vs. normal stroma, but typically needs multiple polarization states. The authors hypothesize that polarization information is fundamentally embedded in single-polarization OCT intensity images because birefringence relates to tissue reflectance and architecture, and therefore could be extracted computationally. The purpose is to develop and validate a deep-learning approach to synthesize PS-OCT contrasts from standard OCT intensity images, potentially enabling PS-like information without specialized hardware.

Prior work shows PS-OCT complements standard OCT and has been used clinically to visualize features and differentiate tissues, including cancer vs. normal, but systems are complex and sensitive to fiber bending artifacts. Conventional PS-OCT computes phase retardation and DOPU by acquiring multiple polarization states. In medical imaging translation tasks, U-Net architectures are common for segmentation but can underperform for complex contrast generation without large datasets. Generative adversarial networks (GANs), particularly pix2pix-style conditional GANs, have achieved superior results in image-to-image translation with smaller datasets and have been applied to virtual histology, digital staining, and synthetic clinical imaging. These findings motivate using a GAN to synthesize PS-OCT contrasts from OCT intensity images.

Image acquisition: PS-OCT images were obtained from fresh human breast tissue with a portable PS-OCT system. A total of 22,072 PS-OCT images (256×512 pixels) were collected from seven subjects with breast cancer and four normal subjects undergoing breast reduction. Each image contains OCT intensity, DOPU, and phase retardation. Human tissue use was IRB-approved; additional tissue came from CHTN. Preprocessing and dataset splitting: Each raw image was cropped to remove dark margins and split into two square images (202×202 pixels). Images were converted from 16-bit to 8-bit. This yielded 44,144 images, split into training:test:validation at 8:1:1. A case-based split (six cancer and four normal cases for training; remaining one cancer and one normal split equally between test and validation) produced SSIMs similar to image-based splitting, indicating robustness to case variance. GAN architecture and training: A modified pix2pix conditional GAN was used, adding SSIM to the loss alongside discriminator loss and L1 distance to improve structural similarity (≥20% SSIM improvement under same training). The generator was a U-Net; the discriminator was a simple three-layer network (deeper discriminators did not improve quality and slowed convergence). Optimization used Adam with learning rate 2×10⁻⁵ and batch size 1 (instance normalization). Training ran for 50 epochs (~16 h) on a Linux workstation (Ubuntu 16.04) with an Nvidia GTX TITAN GPU. Two separate models were trained to synthesize DOPU and phase retardation from OCT intensity. Classification with ResNet-18: To assess utility of synthetic images, the GAN test set (4,414 images) was split 6:3:1 for training:test:validation to train a ResNet-18 classifier for cancer vs. normal. The final fully connected layer was modified to 512×2. Transfer learning was used with ImageNet-pretrained weights; all layers except the final fully connected layer were frozen, improving performance and reducing training time compared to training all layers (which yielded AUC ~0.85). The classifier trained for 24 epochs; the best validation model was evaluated on 350 test images to generate ROC curves and AUCs. t-SNE visualization: The 512-element activation vector from the penultimate layer of the classifier was extracted for each image. t-SNE was performed with Euclidean distance, perplexity 30, and random initialization to visualize distributions of real vs. synthetic PS-OCT images in feature space. Cross-system testing: Models were retrained on PS-OCT images of chicken tissues and applied to OCT-only images from a separate OCT system with similar wavelengths. Imaging was performed on chicken skin and muscle; corresponding sites were identified for comparison to real PS-OCT images, noting local discrepancies due to micron-level co-registration challenges.

- Visual similarity: Synthetic DOPU and phase retardation overlays resembled real PS-OCT images across adipose, stroma, and tumor tissues. - SSIM: After masking pixels below an intensity threshold, SSIM between synthetic and real images was 0.8531 ± 0.0699 for DOPU and 0.6659 ± 0.0517 for phase retardation. Lower SSIM for phase retardation was attributed to higher intrinsic noise/lower SNR. - Classification performance: ResNet-18 classifiers trained on synthetic vs. real PS-OCT yielded similar ROC curves and AUCs. DOPU: AUC 0.979 (synthetic) vs. 0.994 (real), difference 0.015. Phase retardation: AUC 0.952 (synthetic) vs. 0.975 (real), difference 0.023. Despite lower SSIM for phase retardation, classification performance remained high and comparable. - t-SNE: Distributions of synthetic and real images in the classifier feature space were similar for both DOPU and phase retardation, supporting interchangeability for downstream tasks. - Cross-system generalization: After retraining on chicken tissues, GANs produced reasonable synthetic PS-OCT contrasts from OCT-only images acquired on a different OCT system, with good large-scale correlation to real PS-OCT images at matched sites, despite local discrepancies due to alignment. - Robustness to split strategy: Case-based data splitting yielded SSIMs (phase retardation 0.65 ± 0.049; DOPU 0.87 ± 0.53) close to image-based splitting results, indicating robustness to case variance.

The study demonstrates that polarization-related tissue information is embedded within standard single-polarization OCT intensity images and can be computationally extracted using a GAN to synthesize PS-OCT contrasts (DOPU and phase retardation). Although phase retardation synthesis shows lower SSIM due to intrinsic low SNR and stochastic noise, classifiers trained on synthetic images achieve AUCs comparable to those trained on real PS-OCT images, indicating that essential tissue features are preserved and that differences in noise patterns have minimal impact on diagnostic tasks. t-SNE analyses further confirm that synthetic and real images occupy similar regions in feature space. Application to images from a separate OCT-only system (after retraining on relevant tissue) shows feasible cross-system use. These results support the potential for computational PS-OCT to reduce hardware complexity and cost while maintaining diagnostic utility.

This work introduces a deep-learning framework that synthesizes PS-OCT contrasts (DOPU and phase retardation) from standard OCT intensity images using a modified pix2pix GAN. Validations via SSIM, classification AUCs (synthetic vs. real), feature-space visualization, and cross-system testing indicate that synthetic PS-OCT images can serve as effective substitutes for hardware-acquired PS-OCT in cancer/normal classification and general PS-OCT image analysis. The approach can lower cost and complexity by obviating specialized PS hardware. Future extensions could include broader tissue types, improved modeling of phase-retardation noise characteristics, and expanded testing across diverse OCT platforms and clinical settings.

- Phase retardation synthesis exhibited lower SSIM due to intrinsic low SNR and random noise patterns, which the GAN struggled to learn precisely. - Cross-system comparisons had local discrepancies arising from challenges in co-registering the same tissue locations across different imaging systems at micron-level precision. - The primary human dataset was limited to breast tissues from a modest number of subjects (seven cancer, four normal), which may constrain generalizability to other tissues and populations. - Cross-system application required retraining on tissue from the new domain (chicken), indicating domain adaptation may be needed for different tissues/systems.