Translation Identifiability-Guided Unsupervised Cross-Platform Super-Resolution for OCT Images

Optical Coherence Tomography (OCT) is a non-invasive technique for obtaining detailed, cross-sectional images of coronary arteries. However, cost-effective OCT systems produce only low-resolution (LR) images. Unsupervised OCT super-resolution (OCT-SR) presents a cost-effective solution, eliminating the need for high-resolution (HR) systems or co-registered LR-HR image pairs. Existing unsupervised OCT-SR methods formulate the SR task as an image-to-image translation problem, and use CycleGAN as their backbone. However, CycleGAN is known to lack translation identifiability that can result in incorrect SR results. Existing methods often empirically combat this issue by using multiple regularization terms to incorporate expert-annotated side information, resulting in complicated learning losses and extensive annotations. This work proposes a translation identifiability-guided framework based on recent advances in unsupervised domain translation. Employing a diversified distribution matching module, our approach guarantees OCT translation identifiability under reasonable conditions using a simple and succinct learning loss. Numerical results indicate that our framework matches or surpasses the state-of-the-art (SOTA) baseline's performance while requiring considerably fewer resources, e.g., annotations, computation time, and memory.