Fusion of Deep Convolutional Neural Networks for Microaneurysm Detection in Color Fundus Images

Microaneurysms (MAs) are common signsof several diseases, appearing as small circular darkish spots in color fundus images. The presence of even a single MA may suggest diseases (e.g. diabetic retinopathy), thus, their reliable recognition is a critical issue in both human clinical practice and computer-aided systems. As for their automatic recognition, deep learning techniques became very popular in the recent years. In this paper, we also apply such deep convolutional neural network (DCNN) based techniques; however, we organize them into a supernetwork with a fusionbased approach. The combination of the member DCNNs is achieved with interconnecting them in a joint fully-connected layer. The advantage of the method is that this large architecture can be trained as a single neural network, and thus, the member DCNNs are also trained with taking the predictions of the other members into consideration. The competitiveness of our approach is also validated with experimental studies, where the ensemble-based system outperformed each member DCNN. As a primary application domain with strong clinical motivation, the methodology was tested for image-level classification. More specifically, a retinal image is divided into subimages to provide the required inputs for the DCNN-based architecture, and the whole image is labeled as a positive case, if the presence of MA is predicted in any of the subimages. Additionally, we also demonstrate how our architecture can be trained to accurately localize MAs with training only the local neighborhoods of the lesions; empirical tests showing solid performance are also enclosed.