Diabetic retinopathy detection with texture features
Ristilä, Tiina (2017)
2017
Sähkötekniikka
Tieto- ja sähkötekniikan tiedekunta - Faculty of Computing and Electrical Engineering
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2017-08-16
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201708241719
https://urn.fi/URN:NBN:fi:tty-201708241719
Tiivistelmä
Diabetic retinopathy is one of the leading causes of visual impairment and blindness in the world and the prevalence keeps increasing. It is a vascular disorder of the retina and a symptom of diabetes mellitus. The health of the retina is studied with non-invasive retinal imaging. However, the analysis of the retinal images is laborious, subjective and the number of images to be reviewed is increasing.
In this master’s thesis, a computer-aided detection system for diabetic retinopathy, microaneurysms and small hemorrhages was designed and implemented. The purpose of this study was to find out, are texture features able to produce descriptive and efficient information for the retinal image classification and is the implemented system accurate. The process included image preprocessing, extraction of 21 texture features, feature selection and classification with a support vector machine. The retinal image datasets that were used for the testing were Messidor, DIARETDB1 and e-ophta.
The texture features were not successful when classifying the retinal images into diabetic retinopathy or normal. The best average accuracy was 69 % with 72 % average sensitivity and 66 % average specificity. The texture features are not that descriptive as global features with a whole retinal image. Additionally, the varying size of the images and variation within a class affected the classification. The second experiment studied the classification of images into microaneurysm or normal by dividing the retinal images into blocks. The texture features were successful when the images were divided into small blocks of size 50*50. The best average accuracy was 96 % with 96 % average sensitivity and 96 % average specificity. The texture features are more descriptive in the local setting since then they can extract finer details.
To ease the clinical workflow of ophthalmologists and other experts, the computer-aided detection system can lower the manual labor and make retinal image analysis more efficient, accurate and precise. To develop the systems further, an optic disc and image quality detectors are needed.
In this master’s thesis, a computer-aided detection system for diabetic retinopathy, microaneurysms and small hemorrhages was designed and implemented. The purpose of this study was to find out, are texture features able to produce descriptive and efficient information for the retinal image classification and is the implemented system accurate. The process included image preprocessing, extraction of 21 texture features, feature selection and classification with a support vector machine. The retinal image datasets that were used for the testing were Messidor, DIARETDB1 and e-ophta.
The texture features were not successful when classifying the retinal images into diabetic retinopathy or normal. The best average accuracy was 69 % with 72 % average sensitivity and 66 % average specificity. The texture features are not that descriptive as global features with a whole retinal image. Additionally, the varying size of the images and variation within a class affected the classification. The second experiment studied the classification of images into microaneurysm or normal by dividing the retinal images into blocks. The texture features were successful when the images were divided into small blocks of size 50*50. The best average accuracy was 96 % with 96 % average sensitivity and 96 % average specificity. The texture features are more descriptive in the local setting since then they can extract finer details.
To ease the clinical workflow of ophthalmologists and other experts, the computer-aided detection system can lower the manual labor and make retinal image analysis more efficient, accurate and precise. To develop the systems further, an optic disc and image quality detectors are needed.