Abstract
Purpose: :
Excellent image quality is an important requirement to gain a reliable and solid diagnosis from retinal photos. But, there is no established method to automatically measure image quality of fundus photos. We propose an approach to compute an automated quality index for retinal fundus images that classifies statistical features.
Methods: :
From one single fundus image, the algorithm computes three feature types. First, a clustering of the pixel values is performed which finds five image structures in an unsupervised way. The cluster sizes as well as the inter cluster differences are taken as features. Second, two sharpness features are computed by performing a blind signal to noise ratio estimation of the gradient image. Third, three Haralick features which describe global image statistics are taken into account. The feature types are combined and a Support Vector Machine classifier is used to compute a probabilistic output of image quality. Our approach does not require any segmentation of image structures. Compared to other quality measures it is a non-reference quality score based on single images.There were 302 fundus images available (Kowa camera, 22.5° FOV) which were rated by three observers based on four quality criteria (illumination homogeneity, recognizability of the papilla, the vessels, and the background). We used this set for evaluation. Based on the mean score of the observers the images were divided in two classes (good and not good). The automated classification performance was measured in a 10-fold cross validation.
Results: :
We reached an accuracy of 90.3% to find the good images (sensitivity 95.1%, specificity 70.9%). The algorithm was compared to the existing retina image quality method, called Image Structure Clustering (ISC), which was reimplemented and applied on our test set. ISC gained an accuracy of 69.5% (sensitivity 69.0%, specificity 70.0%). The new approach clearly shows a better result on our data set.
Conclusions: :
The proposed method provides a reliable fundus photo rating in terms of general image quality. The algorithm will be used to check the photos at the point of acquisition and give feedback to the operator. It can further be used as an objective measure for selecting high quality images for a screening study or a further automated processing.
Keywords: image processing • imaging/image analysis: non-clinical • computational modeling