Research On Retinal Image Quality Assessment Based On Random Forest

The quality assessment of retinal fundus image is the important problem for the analysis of retinal image. Their quality affects the accuracy and reliability of the diagnosis result and retinal fundus images with high quality is the premise for accurate diagnosis of the disease. However, there are many factors affecting the quality of the fundus images. Firstly, the image may be generates various kinds of distortion during the process of acquisition and transmission. The distortion can lead to the changes of the image quality and distinguished by the human visual system. Furthermore, patient’s head or eye movement, blinking, poorly dilated, small pupils and media opacity all will affect the quality of retinal image. All these factors will lead to the presence of blur, noise, poor illumination and other issues in an image. Now the main route for ophthalmic clinical examination is the subjective judgment by the physicians through analysis of eyes or fundus images.. As a typical image for medical diagnosis, the retinal images contain a complex internal features and a variety of shapes. Putting these images as the research object, and taking some key problems in algorithm for the automatic analysis for further study, is of important theoretical significance and extensive applicability in the medical field. As a result, the retinal image quality assessment method in computer aided diagnosis of fundus image is especially important.The existing methods can be mainly divided into two categories:(i) classification based approaches and (ii) quality metrics based methods. Quality metrics based methods can be further classified into a) segmentation based approaches and b) histogram based approaches. There are many questions with the existing methods. Classification based approaches generally extract features from retinal images and use a classifier to obtain the quality score. These kinds of methods usually take feature extraction on full images which contain both useful information and nonessential information. Other methods use image features information inaccurately and insufficiently in the process of evaluation, and it is time-consuming and usually the segmentation result is inaccurate.Based on these backgrounds, a no-reference quality assessment method for retinal image based on Random Forest is introduced in this paper. It mainly focuses on the blur and noise of an image. The method combines supervised classification method and unsupervised evaluation method, targets for feature extraction and classification, and finally obtains good results in the evaluation of the retinal fundus images.First, the images are enhanced by curvelet transformation to reduce the influence of noise when choosing the anisotropic patches. Second, anisotropic patches are selected on the enhanced images and then extracted on the original images. Then, nine features are adopted on the original green channel at the extracted anisotropic patches, which are effective with regard to blur and noise. Subsequently, random forest is used to classify the selected patches in an image. Finally, the image quality is obtained by the voting of classifying result. The algorithm is demonstrated to be effective to measure quality of retina images. It proved to be effectiveness for this method on the aspect of fundus retinal image quality assessment.