Joint Segmentation of FAZ and RVs in OCTA Images With Auxiliary 3-D Image Projection Learning

In the field of ophthalmic clinical practice, the quantitative and qualitative analyses of the foveal avascular zone (FAZ) and retinal vessels (RVs) in optical coherence tomography angiography (OCTA) images hold significant importance for assisting the diagnosis and analysis of eye diseases. However, the segmentation of FAZ and RV is typically based on 2-D images and is naturally separated. In this article, we propose a joint segmentation framework called PL-Joint-Seg, which incorporates 3-D image projection learning (PL) and forms a dual-path structure: 1) in the 3-D image projection path, volume features are extracted using an image projection network with an encoder–decoder-like structure and 2) in the 2-D image segmentation path, 3-D projection features are fused with 2-D encoding features to establish consistency and complementarity (CsCp) between features of different dimensions, leveraging both 3-D structural information and 2-D spatial information for joint segmentation. Furthermore, feature alignment decoding blocks and a multiscale soft fusion module (MSFM) are further inserted into the 2-D image segmentation path to adapt to the differences in the morphological characteristics of FAZ and RV. Finally, we evaluate our method using the publicly available OCTA-500 dataset. The proposed PL-Joint-Seg achieves Dice scores of 0.9110/0.8982 and 0.9848/0.9165 for FAZ/RV segmentation on the OCTA_6M and OCTA_3M datasets, respectively, which demonstrate that adding a 3-D projection path plays an important auxiliary role in guiding 2-D image segmentation, while PL-Joint-Seg exhibits good performance in noise robustness and computational efficiency.