CD-Net: A Cross-Dimensional Feature Fusion Network for Retinal Vessel Segmentation Based on OCTA Images

Extracting retinal vessels from Optical Coherence Tomography Angiography (OCTA) images is crucial for computer-aided diagnosis and treatment of various ocular diseases. However, existing methods fail to fully leverage the retinal vessel features across different dimensions and lack attention to the topological morphology of retinal vessels. To address these issues, a novel OCTA retinal vessel segmentation network based on cross-dimensional feature fusion, named CD-Net, is proposed. This network comprises a cross-dimensional learning path and a planar learning path, jointly utilizing 3D volumetric data and 2D projection images as input.Firstly, considering the complementarity and specificity of data across different dimensions, the cross-dimensional learning block (CDLB) and vessel topology enhancement module (VTEM) are designed within the cross-dimensional learning path. These components aim to capture finer features of microvasculature and highlight the topological morphology of planar vessels. Secondly, to compress 3D features into 2D while retaining richer vessel feature information, a stride compression module (SCM) is employed to progressively reduce the dimensionality of features to 2D. Thirdly, features from the two encoding branches are effectively fused at both the channel and spatial levels to obtain cross-dimensional vessel features. Finally, in the planar learning path, the complementary features generated by each cross-dimensional learning block are merged into the corresponding planar encoding block, further integrating cross-dimensional vessel features to achieve the final 2D segmentation result. Experimental results on the OCTA-500 dataset demonstrate that CD-Net achieves a Dice similarity coefficient of 90.39%, outperforming the comparison methods mentioned in the text.