Automatic Detection of the Optic Disc in Retinal Images
This project is an attempt to find the optimal way of detecting the optic disc (OD) within a retinal image by experimenting with and evaluating a variety of techniques so that the best method can eventually be used to help in the diagnosis of glaucoma. Currently detection is dependent upon either a combination of tonometry, perimetry or Heidelberg retinal tomography (HRT).
Retinal images, of a standard currently used by ophthalmologists to diagnose glaucoma, have been made available. There are 196 in total, 110 are from healthy eyes and 86 are from glaucoma patients. They will provide the means to generate and test a software solution using C++ and OpenCV. Three experiments will be attempted using Hough Transform, Active Contour Models (Snakes) and Active Appearance Model in conjunction with Thresholding, the Sobel operator, the Canny operator and morphological filtering.
1) The first experiment to locate the OD will use the Hough transform. The points that will be provided for the edge will be found using thresholding, the Sobel operator and the Canny operator individually and the results of each compared.
2) The second experiment to locate the OD will use multiple snakes and expand upon the previous method. The snakes will be positioned in the image by using the Hough transform. The radius of the Hough transform result will be made slightly larger and the snake nodes positioned there so that it can tighten around the OD. A second snake will be positioned within the OD and will expand so an average border can be created from the two results. Depending on the initial success rates of the snakes a pre-processing stage may be introduced which will remove the blood vessels from the image using morphological filtering.
3) The an active appearance model will be built from a test set of images and used to attempt to match the optic disc border within the remaining images. The parameters of the model when it has fit to the image will define the border.
An automatic evaluation system has been designed which will compare the results of the optic disc detection to an outline of the optic disc border provided by a human expert. A score between 0 and 1 will be calculated based on the difference between the two borders where 1 represents a perfect match and 0 a complete failure. The success of the HRT system will be the benchmark for the results.
With successful detection of the OD to a high level of accuracy (near 1) the system will have solved the initial problem of glaucoma detection. If the results are significantly below HRT accuracy the project will bring to light the failings of the techniques within this field and allow future research to be directed elsewhere. The last section contains extensions of this project outlining the work intended to be completed in the future and an overview of who the work will benefit.