Retinal blood vessel segmentation via geodesic methods in image processing

Automated detection of retinal blood vessels via image processing techniques is an imoprtant task for diagnosing the miscellaneous eye diseases. In this study, the retinal images and the vessels were respectively handled as 2D surfaces embedded into the 3D space and valley stripes overlaying on those surfaces. Firstly, the discrete version of the mean surface curvature matrix of the smoothened retinal image was calculated and then it was observed that the sink points on that matrix correspond to the the vessel borders and the vessel regions resemble to the hills. Then, a constraint map matrix which specifies the approximate borders of the vessel like regions by exploiting that surface curvature characteristic was obtained. Afterwards, a connected vessel component which contains 3D geodesic distance map information was segmented using the Fast Marching Algortihm and at least one seed point inside the constraint map matrix. Later, the main part of the proposed semi automated system was converted to the full automated version by using the seed points obtained from the histogram of the maximum normal surface curvature matrix. However, it was observed that the proposed method misses the pixels on the borders of the main wide vessels and then its performance was enhanced via a kernel based submethod benefiting from the neighbourhood statistics of the pixels. Additionally, two new lesion removal algorithms were also applied in this study. Lastly, small holes on the vessel segments were filled via morphological operations in order to eliminate the side effect which is known as the central vessel reflex occuring because of the lighting technique of the imaging device. This study was tested on DRIVE, STARE and CHASE_DB1 data sets and it was observed that it acquired admissible results in terms of sensitivity, specificity, accuracy and execution time with respect to the recent studies in the literature.