In this research work, Random Forest classifier is used for determining different stages of retinal abnormalities due to DR using machine learning techniques. Being an ensemble classifier, Random Forest constructs several decision trees at training time and generates the classification for each tree. In this research work, a dataset containing several retinal images having abnormalities is formed. The images are collected from various sources like DIARETDBO and DIARETDB1 . A set of nine features including three statistical features and six texture-based features are selected for the machine learning. For each input image, the feature values are calculated, and thus, a dataset is formed for 69 images. In Weka 3.7 [11], the dataset is supplied as input to the Random Forest classifier. Performance measures like accuracy, sensitivity, and specificity are calculated depending on the classification result. The accuracy of HAM and MA classes is 100% each as the feature size distinctly separates these two classes. The size of HAM is considered to be medium to large while MA is very small. In future, a collection of large database with more features can be added to the feature set, and the number of images can be increased to get more complex training set for the classifier. The average accuracy is 99.275% which is promising.
展开▼
