Detection of microaneurysms in retinal images using an ensemble classifier

This paper introduces, and reports on the performance of, a novel combination of algorithms for automated microaneurysm (MA) detection in retinal images. The presence of MAs in retinal images is a pathognomonic sign of Diabetic Retinopathy (DR) which is one of the leading causes of blindness amongst the working age population. An extensive survey of the literature is presented and current techniques in the field are summarised. The proposed technique first detects an initial set of candidates using a Gaussian Matched Filter and then classifies this set to reduce the number of false positives. A Tree Ensemble classifier is used with a set of 70 features (the most commons features in the literature). A new set of 32 MA groundtruth images (with a total of 256 labelled MAs) based on images from the MESSIDOR dataset is introduced as a public dataset for benchmarking MA detection algorithms. We evaluate our algorithm on this dataset as well as another public dataset (DIARETDB2) and compare it against the best available alternative. Results show that the proposed classifier is superior in terms of eliminating false positive MA detection from the initial set of candidates. The proposed method achieves an ROC score of 0.415 compared to 0.2636 achieved by the best available technique. Furthermore, results show that the classifier model maintains consistent performance across datasets, illustrating the generalisability of the classifier and that overfitting does not occur.