Time-step encoded high-frequency enhanced diffusion model for OCT retinal image denoising

Optical coherence tomography (OCT) is a widely used imaging technique in ophthalmology, capable of non-invasive, high-resolution imaging of retinal tissues. However, OCT images are often degraded by speckle noise, resulting in poor image quality. Deep learning-based denoising models have become the mainstream approach, but existing methods tend to oversmooth images and lose high-frequency details, making it difficult to recover the true retinal structure. This paper proposes a high-frequency enhanced diffusion model based on the cold diffusion framework, named THFN-OCT (time-enhanced high-frequency network for OCT denoising). The model decouples frequency-domain information and processes it separately, with cross-domain connections to preserve high-frequency details while ensuring denoising performance. In addition, considering the different impacts of each diffusion timestep on frequency components, we design a timestep-aware attention module that uses the timestep t to guide the reconstruction. Experiments on two public OCT retinal denoising datasets and one private dataset show that the proposed method outperforms existing denoising algorithms.