UC San Diego

Multi-modal retinal images provide complementary anatomical information at various resolutions, color wavelengths, and fields of view. Aligning multi-modal images will establish a comprehensive view of the retina and benefit the screening and diagnosis of eye diseases. However, the inconsistent anatomical patterns across modalities create outliers in feature matching, and the lack of retinal boundaries may also fool the intensity-based alignment metrics, both of which will influence the alignment qualities. Besides, the varying distortion levels across Ultra-Widefield (UWF) and Narrow-Angle (NA) images, due to different camera parameters, will cause large alignment errors in global transformation.

In addressing the issue of inconsistent patterns, we use retinal vasculature as a common signal for alignment. First, we build a two-step coarse-to-fine registration pipeline fully based on deep neural networks. The coarse alignment step estimates a global transformation via vessel segmentation, feature detection and description, and outlier rejection. While the fine alignment step corrects the remaining misalignment through deformable registration. In addition, we propose an unsupervised learning scheme based on style transfer to jointly train the networks for vessel segmentation and deformable registration. Finally, we also introduce Monogenical Phase signal as an alternative guidance in training the deformable registration network.

Then, to deal with the issue of various distortion levels across UWF and NA modalities, we propose a distortion correction function to create images with similar distortion levels. Based on the assumptions of spherical eyeball shape and fixed UWF camera pose, the function reprojects the UWF pixels by an estimated correction camera with similar parameters as the NA camera. Besides, we incorporate the function into the coarse alignment networks which will simultaneously optimize the correction camera pose and refine the global alignment results.

Moreover, to further reduce misalignment from the UWF-to-NA global registration, we estimate a 3D dense scene for the UWF pixels to represent a more flexible eyeball shape. Both the scene and the NA camera parameters are iteratively optimized to reduce the alignment error between the 3D-to-2D reprojected images and the original ones, which is also concatenated with the coarse alignment networks with distortion correction function.