Fovea localization by blood vessel vector in abnormal fundus images

Highlights：

• BVV models the global feature of the vasculature including the minute and caterpillar vessels aside from the major vessels. As an unsupervised technique, the BVV model is distinct from the parabola which is determined only by a few main blood vessels and sensitive to the noise.

• The proposed method does not need the physical coordinate of the anatomic structures in fundus images. The BVV model only depends on relative position of anatomic structures without involving the coordinate transformation, hence the calculation is more convenient than the parabola model.

• The proposed method can locate the fovea robustly in a small searching region in fundus images with dark lesions. The global statistic characteristics of the blood vessels makes the BVV accurately and robustly estimate the retinal raphe, thus the heuristic searching can be accomplished in a small region.

摘要

•A new position-related model based on the blood vessel vectors (BVVs) is proposed to locate fovea in abnormal fundus images. The model-driven deep learning technique is developed to train the OD to remove the distraction of big bright lesions. After detecting blood vessels and OD, the fovea can be estimated accurately by summating BVVs.•BVV models the global feature of the vasculature including the minute and caterpillar vessels aside from the major vessels. As an unsupervised technique, the BVV model is distinct from the parabola which is determined only by a few main blood vessels and sensitive to the noise.•The proposed method does not need the physical coordinate of the anatomic structures in fundus images. The BVV model only depends on relative position of anatomic structures without involving the coordinate transformation, hence the calculation is more convenient than the parabola model.•The proposed method can locate the fovea robustly in a small searching region in fundus images with dark lesions. The global statistic characteristics of the blood vessels makes the BVV accurately and robustly estimate the retinal raphe, thus the heuristic searching can be accomplished in a small region.