CAMNet: A combined attention-driven deep learning framework for OCT image denoising

Optical coherence tomography (OCT) speckle noise caused by low coherent light interference can significantly degrade image quality and obscure fine structural details in OCT images, thus affecting diagnostic accuracy. In this work, we demonstrate the Combined Attention Mechanism Network (CAMNet) for speckle noise reduction, which integrates spatial attention and channel attention modules. Specifically, CAMNet utilizes a Mixed Attention Block (MAB) constructed with a Residual-on-Residual architecture to capture deep feature representations. Within this framework, the Spatial Attention Module (SAM) exploits dual-pooling strategies to generate attention maps that spatially localize informative regions, while the parallel Channel Attention Module (CAM) automatically learns channel-wise dependencies to assign differentiated weight coefficients to key feature maps. This synergistic interaction enables the network to overcome the local receptive field limitations of traditional CNNs by effectively aggregating global contextual information with local structural features. Extensive experimental evaluations demonstrate that CAMNet outperforms traditional methods and deep learning methods in speckle noise reduction and structural fidelity preservation. Quantitatively, the proposed CAMNet yields an average PSNR of 30.3154 and SSIM of 0.7075 on this dataset; Qualitatively, the images processed by CAMNet reduce speckle noise as well as preserve the layered structure of the images. © 2026 Elsevier Ltd