Automatic fundus image classification for computer-aided diagonsis.

With the advances of computer technology, more and more computer-aided diagnosis (CAD) systems have been developed to provide the "second opinion". This paper reports an automatic fundus image classification technique that is designed to screen out the severely degraded fundus images that cannot be processed by traditional CAD systems. The proposed technique classifies fundus images based on the image range property. In particular, it first calculates a number of range images from a fundus image at different resolutions. A feature vector is then constructed based on the histogram of the calculated range images. Finally, fundus images can be classified by a linear discriminant classifier that is built by learning from a large number of normal and abnormal training fundus images. Experiments over 644 fundus images of different qualities show that the classification accuracy of the proposed technique reaches above 96%.

Duke Scholars

Author Tien Yin Wong Ophthalmology