Cross-Platform Super-Resolution for Human Coronary Oct Imaging Using Deep Learning

Optical coherence tomography (OCT) has emerged as a preferred imaging method for assessing plaques before stenting and understanding blood vessel responses to intervention. However, the current image resolution still limits the effective capture of crucial intravascular elements. Although deep learning-based super-resolution techniques, relying on high-resolution (HR) and low-resolution (LR) pairs, hold promise in enhancing image resolution, existing methods primarily employ HR and LR images from the same imaging platform to demonstrate the potential of deep learning. This approach is impractical in real imaging scenarios where the HR image can not be obtained from a LR imaging platform. In this paper, we present a cross-platform deep learning framework that leverages unpaired cross-platform datasets. The HR training dataset is sourced from a high-end, high-cost OCT system, while the LR training dataset originates from a low-end, low-cost OCT system. Improving a Cycle Generative Adversarial Network with a specialized focus on coronary image structure, our experiments indicate that the new network generates super-resolved images from any LR image, demonstrating image quality comparable to OCT images acquired by HR systems.