Polarization-sensitive optical coherence tomography (PS-OCT) is a valuable biomedical imaging technique that provides additional diagnostic information by analyzing the polarization of scattered light, revealing tissue birefringence. This birefringence is linked to the structure and organization of tissues like collagen and muscle fibers, enabling better differentiation of tissues like cancer and connective tissues, which appear similar in standard OCT images. PS-OCT has shown promise in clinical applications, particularly in differentiating cancer from normal tissue. However, traditional PS-OCT systems are complex and expensive due to the requirement of multiple detectors and polarization-manipulating components. Handheld probe integration is also challenging due to polarization changes in the fiber optic cable. This complexity limits widespread adoption, especially in intraoperative settings where real-time imaging of tumor margins is crucial. The researchers hypothesize that the polarization information in PS-OCT is inherently encoded within the OCT intensity image, making it possible to extract this information computationally. Deep learning, specifically generative adversarial networks (GANs), have proven effective in image-to-image translation tasks. Therefore, they propose using a GAN to synthesize PS-OCT images from standard OCT intensity images, eliminating the need for expensive and complex hardware.

The paper reviews existing literature on PS-OCT, highlighting its advantages and limitations in clinical applications. It also discusses previous applications of deep learning, specifically GANs, in various image translation tasks in biomedical imaging, such as virtual histology, digital phase staining, and synthetic clinical imaging. The authors acknowledge the limitations of U-Nets in generating complex image contrasts, particularly with limited datasets, and thus opt for GANs, whose architecture with both generator and discriminator networks is better suited for this task.

The researchers used a GAN, modified from the pix2pix GAN, to synthesize PS-OCT images. The GAN was trained on paired OCT intensity and PS-OCT images acquired from fresh human breast tissue samples (cancerous and normal). The PS-OCT images contained two common polarization contrast metrics: degree of polarization uniformity (DOPU) and phase retardation. The dataset consisted of 22,072 PS-OCT images (256 × 512 pixels) from 11 subjects (7 cancer, 4 normal). Images were preprocessed (cropped, converted to 8-bit) before training. The dataset was split into training, testing, and validation sets (8:1:1 ratio). A U-Net served as the generator, and a three-layer discriminator network was used. The loss function included the discriminator loss, L1 distance, and SSIM to improve image quality. The Adam optimizer with a learning rate of 2 × 10⁻⁵ and a batch size of 1 was used. The training process, performed on a Linux machine with GPU acceleration, took about 16 hours for 50 epochs. To validate the synthetic PS-OCT images, a ResNet-18-based image classifier was trained separately on both real and synthetic PS-OCT images for cancer/normal classification. The test dataset (4,414 images) for the GAN was further divided for training, testing, and validation of the classifier. Transfer learning was employed to improve training efficiency and accuracy by initializing the network with pre-trained weights from ImageNet and fixing all weights except the last fully connected layer. The classifier performance was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC). t-SNE analysis was also performed to visualize the distribution of real and synthetic PS-OCT images in the feature space. Finally, the trained GAN was applied to OCT images acquired from a separate standard OCT system (without PS-OCT capabilities) using fresh chicken tissue samples to further assess its generalizability.

The SSIM values between real and synthetic PS-OCT images were 0.8531 ± 0.0699 for DOPU and 0.6659 ± 0.0517 for phase retardation. The lower SSIM for phase retardation was attributed to higher noise levels in these images, impacting the GAN's ability to accurately synthesize the noise pattern. However, the cancer/normal classification results showed similar performance between classifiers trained on real and synthetic PS-OCT images, with AUC values approaching 1 (0.979 for synthetic DOPU, 0.994 for real DOPU; 0.952 for synthetic phase retardation, 0.975 for real phase retardation). The similarity in AUC values suggests that synthetic PS-OCT images are effective substitutes for real images in classification tasks. The t-SNE analysis visually confirmed the similarity in the distribution of real and synthetic PS-OCT images in the feature space. Furthermore, the application of the trained GAN to OCT images from a separate system produced synthetic PS-OCT images that correlated well with real PS-OCT images obtained from the same tissue sites, demonstrating the model's generalizability.

The findings demonstrate the successful synthesis of PS-OCT images from standard OCT intensity images using deep learning. The high classification accuracy achieved using the synthetic PS-OCT images shows their potential to replace real PS-OCT images in various applications, significantly reducing the cost and complexity of PS-OCT imaging. The good performance of the model even with the lower SSIM for phase retardation images highlights the classifier's ability to focus on meaningful tissue features, despite the differences in noise patterns. The ability to apply the trained model to images from a different OCT system further supports the generalizability of the method. This approach could pave the way for wider adoption of PS-OCT technology in clinical settings, particularly in intraoperative applications.

This study successfully demonstrated a deep-learning-based computational method for synthesizing PS-OCT images from standard OCT images, eliminating the need for expensive and complex hardware. The high accuracy of the synthesized images in cancer/normal classification tasks and their successful application to a different imaging system highlight the method's potential to broaden the use of PS-OCT technology. Future research could focus on improving the synthesis of noisy features like phase retardation, exploring different deep learning architectures, and validating the approach in larger clinical studies.

The study's limitations include the relatively small sample size used for training the GAN and classifier, although sufficient images were included. The generalization of the method to different tissue types and imaging systems needs further investigation. The noise level present in the phase retardation images might limit its application. The performance might vary when applied to tissues with vastly different optical properties compared to the breast tissues used in this study.