Purchase this article with an account.
C. Agurto Rios, V. Murray, S. Barriga, M. S. Pattichis, W. C. Bauman, Jr., P. Soliz; Automatic Classification of Diabetic Retinopathy Photographs Using Am-Fm. Invest. Ophthalmol. Vis. Sci. 2010;51(13):1795.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To validate a classification algorithm for the detection of Diabetic Retinopathy (DR) in color fundus digital photographs using amplitude modulation-frequency modulation (AM-FM) as feature extraction technique and Partial Least Squares (PLS) as a classification method.
A DR screening algorithm has been tested on 2 different databases. The first dataset consisted of 1200 images from MESSIDOR, and 500 images from the University of Iowa department of Ophthalmology provided by Dr. Michael Abramoff. The images were categorized by ophthalmologists into 4 levels of DR severity. Multiscale AM-FM, a mathematical technique that extracts features from images in different frequency bands, is applied to each image. Images are further subdivided in regions of interest (ROIs). A total of 39 features are extracted for each region, corresponding to the 3 estimates produced by AM-FM and 13 combinations of bandpass filters. An unsupervised clustering method (k-means) is used to group similarities in the ROIs prior to computer classification. Testing is done using the cross validation method, where the training and testing sets of images are chosen randomly from our dataset.
The MESSIDOR database is divided in three sets of 400 images each. The results obtained for each of the 3 sets are: AUC1=0.86, AUC2=0.84, and AUC3=0.85, corresponding best sensitivity and specificity values were 98%/67%, 92%/66%, and 95%/70%. For the U of Iowa database, we obtained an AUC of 0.82, with 91%/65% sensitivity/specificity. An additional test was performed for the classification of images containing Sight Threatening DR. An AUC=0.98 and sensitivity/specificity of 100%/88% was obtained for this case.
The classification results obtained with our algorithm are comparable to the published results in systems of detection of DR. As opposed to other methods, ours is a top-down approach not requiring manual segmentation of lesions. In addition, the feature extraction using AM-FM is proven to be robust since different sizes of images and different places of acquisition for the database are used in this implementation without significant variation in the results.
This PDF is available to Subscribers Only