Biologically-Inspired Supervised Vasculature Segmentation in SLO Retinal Fundus Images

We propose a novel Brain-Inspired Multi-Scales and Multi-Orientations (BIMSO) segmentation technique for the retinal images taken with laser ophthalmoscope (SLO) imaging cameras. Conventional retinal segmentation methods have been designed mainly for color RGB images and they often fail in segmenting the SLO images because of the presence of noise in these images. We suppress the noise and enhance the blood vessels by lifting the 2D image to a joint space of positions and orientations (SE(2)) using the directional anisotropic wavelets. Then a neural network classifier is trained and tested using several features including the intensity of pixels, filter response to the wavelet and multi-scale left-invariant Gaussian derivatives jet in SE(2). BIMSO is robust against noise, non-uniform luminosity and contrast variability. In addition to preserving the connections, it has higher sensitivity and detects the small vessels better compared to state-of-the-art methods for both RGB and SLO images.