Automatic Algorithms For Quality Improving, Feature Detection And Registration From Retinal Funds Images

Retinal fundus images play an important role in ophthalmology diagnosis. Currently, image processing, analysis and computer vision techniques are widely used in retinal image quality improving, feature detection, image registration and stitching, and other automatic retinal image analysis algorithms, which can help ophthalmologist to diagnosis the retinal disease and other systemic disease quickly and efficiently. Combination the requirements of clinic and automatic analysis algorithms, the research is divided into three parts.1. Image quality improving: A enhancement method for fluorescein retinal image is presented using Gabor filter and morphology which is realized by adjusting the weight coefficient dynamically. An efficient illumination equalization approach based on illumination and reflectance model is proposed. The illumination algorithm is realized by four steps: (1) Separate the original image into the background image and foreground image using multi-scale processing. (2) Estimate the background image by multi-directional line-average method. (3) Equalize the background image through dividing the image by illumination component. (4) Apply gray scale correction to foreground image. The experiments demonstrate the two methods can eliminate the detrimental effect and non-uniform illumination resulted from imperfect imaging conditions while persevering main features well. The performance of the proposed methods outperforms many mainstream enhancement and illumination correction methods.2. Image feature detection: A fast and efficient approach to localize and segment the optic disc in fluorescein retinal images by means of mathematical morphology and the GVF snake model is proposed. The localization algorithm utilizes the similarity in gray intensity between blood vessels and optic disc. After localization, the pixel-level preprocessing is performed to remove blood vessels, where the optic disc boundary is then detected by the GVF snake. The correct rates of disc localization are 96.7% and the boundary detection method yields average sensitivity values of 92%. An efficient and precise retinal vessel centerline extraction approach based on ridge detection is proposed. The method starts by obtaining the centerline candidate using ridge detection. The centerline segment is validated by logic AND operation between the centerline candidate and the summation of the binary image of each scale which is obtained by Otsu dual threshold method and Otsu single threshold method respectively. The experiments demonstrate the method can extract the slim vessels and low contrast vessels well while persevering vessels connectivity well. The average overlap rate of centerline extraction is 83.5%.Image registration: A coarse-to-fine registration scheme for retinal images is presented using the combination of the information of vessel centerline trees, bifurcations and crossovers, and pixel intensity. Multi-resolution based coarse matching algorithm and mutual information based fine matching algorithm are used to obtain the optimal correspondences and accurately estimate the non-linear transformation. Experiments were performed on 160 image pairs from 55 different eyes. The result shows when the overlapping percentage of the image pairs are greater than 30%, the mean of success rate is up to 85% and when the overlapping percentage of the image pairs are greater than 60%, the mean of success rate is up to 95%.