Retinal Image Registration Based on Features of Vessel-Segmented Image

Goal: The registration of retinal images captured at different view angles and different time or by different imaging systems plays a crucial role in diagnosis of ophthalmic disease. However, existing methods which are typically solved by matching feature points suffer from great challenges due to poor image quality, different image modalities and partial overlap. Methods: In this paper, we propose a joint vessel segmentation and speed up robust feature (SURF) based registration method implemented in both monomodal and multimodal retinal images. We adopt Dense U-net and patch-based learning strategy to realize the vessel segmentation of retinal images after our preprocessing, and then we extract the vessel features based on SURF detector to realize registration. The method was performed on two public challenging datasets and was evaluated by root-mean-square-error (RMSE) and Mean Absolute Deviation (MAD). Results: The results show that our method can acquire more key points and more correct matching pairs, and the success rate surpasses 90% in two datasets. The robustness and accuracy of our method outperform four state-of-the-art registration methods. Conclusion and significance: These findings demonstrate that our method can integrate the complementary information from different modality retinal images and improves the efficiency and accuracy of diagnosis.