Automated Quality Assessment Of Color Fundus Images

Fundus images have become an intuitive and standard diagnosis method in recent years.A large number of fundus images are taken out in hospital,which contain good and bad quality images and often need well-trained ophthalmologists or experts to classify.Because there exists problems of a large number of fundus images,time consuming and experts'scarcity,it is important to explore an objective and automatic system to evaluate the quality of fundus images.My goal is to find an automated and objective measurement of the image quality of single retinal fundus photos to allow a stable and reliable medical evaluation.The retinal image quality evaluation system presented here bases on the classification of global and local features that correlate with the human perception of retinal image quality as assessed by eye care specialists.This study is based on 360 typical fundus images,which classified by the ophthalmologist and 180 good ones and 180 bad ones,then take 90 normal quality and 90 bad quality images as training set and the remained images(90 normal and 90 bad quality images)as testing set.The system consists of two main processing phases:features extraction and construct classifier model.The first phase is the overall image features extraction,such as illumination uniformity,brightness,and contrast are measured by global histogram and textural features;the sharpness of local structures,such as optic disc and vasculature network,is measured by gradient and local vessel density.The second phase is fundus images are trained by different classifiers and chosen the best one as the final model.For any given retinal fundus image,by extracting the features and input the model,we can know its quality.The SVM classifier with radial kernel function has the best performance through testing.After putting the testing dataset into the model to calculate the performance,the result is that an area under the ROC curve(AUC)is 0.971 and accuracy is 91.11%.The value of AUC and accuracy is relatively high.In determining the image quality of retinal fundus images,this method can automatically produce reliable and objective results.The combination of local and global features to define the image quality evaluation standard is feasible.