Fundus Image Registration with Binary Morphology Extraction of Feature Points

Fundus image registration is essential for clinical eye disease examination, in which diseases such as diabetic retinopathy or macular degeneration may require a meticulous monitoring. Multiple retinal images may be registered to analyze the evolution of patients over time, to widen the field of view or to enhance the resolution for a detailed examination. At present, feature-based fundus registration methods prevail, nonetheless, some methods have a great feature points density, which may complicate the matching process for some feature descriptors due to the similarity among the points. Furthermore, several methods for vessel structure segmentation have been developed, occasionally employing Hessian matrix features. However, the use of these features, has not been extensively employed for registration purposes. This paper proposes a fundus image registration with binary morphology extraction of feature points that involves the frangi filter for demarcating a region of interest, prioritizing the points distribution over the abundance. Later, medial axis transform and pattern detection are made for obtaining feature points that are characterized by the Fast Retina Keypoint (FREAK) descriptor and matched for computing the transformation matrix. The proposed method is assessed with the Fundus Image Registration Dataset (FIRE). Results suggest that the proposed method can compete to certain extent with some of the former similar approaches in regard to registration error, achieving a 0.5084 for area under the curve.