Microaneurysms Detection in Color Fundus Image with Feature-based Background Suppression

Anneke Annassia Putri Siswadi, Stéphanie Bricq, Fabrice Mériaudeau

Abstract: Diabetes is one of the major causes of blindness. Diabetic retinopathy (DR) is the most frequent complication of diabetes. One in three people suffering from diabetes would develop diabetic retinopathy. However, the risk of vision loss caused by DR could be prevented by having early treatment. Microaneurysm (MA) is a tiny-red-spot lesion that appears in the fundus image as the first symptom of DR. Automatic detection of MAs is growing research due to the increasing need for efficient and accurate detection. However, MAs detection is prone to false-positive detection because of the imbalance data issue. It is necessary to minimize the number of false-positive, while sensitivity should be as high as possible. In this research, we trained and tested the method in individual dataset to evaluate the sensitivity of the method performance. The proposed method requires less number of data to classify the MAs. This is achieved by maximizing the information extracted from the positive data in each process: maximize the output of CLAHE image enhancement technique, extract MAs candidates in unsupervised approach, enrich the object features by novel local background suppression technique, and apply the cascade learning of two identical unit networks. The cascade learning acts as the filter detection that pushes the classifier to focus learning only on the hard cases. The evaluation shows that the proposed method can reduce the number of false positives significantly. The proposed network is trained and tested with E-Ophta and IDRiD datasets individualy and it can reach the highest (individual) sensitivity with 79.2% for 8 FPI in the FROC evaluation metric.