Punctured window based multiscale line detector for efficient segmentation of retinal blood vessels

Changes in the retinal vasculature can help diagnose diseases like diabetes, hypertension, and arteriosclerosis. To enable ophthalmologists to provide an efficient diagnosis and further reduce the cost of treatment, we propose an automated algorithm for the segmentation of retinal vasculature. Line detector is a classic approach for vessel-like structure detection or segmentation but with a fundamental flaw in estimation of the background intensity around a pixel. In this work, we highlight and rectify that issue in the classic line detector, by introducing the idea of punctured windows. This enhances the ability of a line detector to identify minor vessels in low contrast regions. Firstly, the image is denoised using a fractional filter. Then, the line detector with punctured window is used to compute the line responses at multiple scales. The final response is computed as the arithmetic mean of all responses at different scales and the underlying image intensity. Finally, hysteresis thresholding is applied to obtain the segmented vessels. The majority of methods proposed in the literature are evaluated only on DRIVE and STARE datasets, and using the performance metrics that are biased due to the issue of class imbalance. While many other methods fail to be consistent either across different datasets or the performance metrics used. The proposed algorithm is tested on four publically available datasets, namely, RC-SLO, STARE, CHASE_DB1, and DRIVE using several performance metrics that are unaffected by the class imbalance prevalent in vessel classification problems. The proposed technique is comparable with state-of-the-art methods and outperforms many of them.