Cross-domain knowledge integration framework for optical coherence tomography images denoising

Optical Coherence Tomography (OCT) is widely used in ophthalmic diagnosis owing to its high-resolution, non-invasive imaging capabilities. However, OCT images are often degraded by speckle and detector noise, obscuring critical morphological structures and compromising image sharpness. Achieving effective noise suppression while preserving fine-grained anatomical details remains a key challenge for existing methods, especially in retaining subtle pathological features. To address this, we propose a novel Cross-Domain Knowledge Integration Framework (CDKIF) that integrates segmentation and classification networks to guide denoising optimization. Specifically, our method distills explicit boundary priors from a segmentation network and transfers this spatial awareness to the denoising network via targeted parameter sharing, significantly enhancing edge preservation. Moreover, domain-adapted perceptual loss, computed using features from a fine-tuned classification network, specifically reduces the blurring of pathological structures and fine layer details. Finally, we propose a multi-stage pipeline combining Cluster-Weighted Multi-scale (CWM) based spatial regularization and temporal averaging to generate high-fidelity noise-free labels, eliminating paired-data dependency and artifact risks associated with conventional averaging. Extensive experiments demonstrate that our method outperforms existing methods in denoising performance and generalization. Critically, the enhanced preservation of structural details and layer boundaries translates to substantial improvements in downstream retinal layer segmentation, validating the clinical utility of our framework.