Ens5B-UNet for Improved Microaneurysms Segmentation in Retinal Images

Abstract

The appearance of Microaneurysms (MAs) in the retina is one of the earliest clinical signs for the diagnosis of Diabetic Retinopathy (DR) and other sight threatening diseases However, manual detection and segmentation of MAs for diagnosis is tedious, time consuming and prone to misclassification. Hence, accurate automated methods may significantly reduce diagnostic workload of clinician and prevent misdiagnosis. Even though, several single network semantic segmentation models have been proposed in literature, ensemble learning remains insufficiently explored despite its established benefits. We propose a novel ensemble learning architecture, Ens5B-UNet that combines the weights of 5 individual modified U-Net models to create a unified approach for improved microaneurysms segmentation. The model was trained on the IDRiD dataset and evaluated on EOphta _MA dataset. Ens5B-UNet performance is 3.17% (AUPR) higher than the 1<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">st</sup> position in the IEEE ISBI – 2018 (International Symposium on Biomedical Imaging) competition. The proposed model was further compared with state-of-the-art models and it outperformed all in IOU, Dice, AUPR and specificity evaluation metrics with scores of 67.76%, 49.07%, 52.74%, and 99.93% respectively. These results are over the best state-of-the-art models by 19.67%, 10.57%, 3.43%, and 6.65% respectively in IOU. Evaluation on E-Ophta_MA shows remarkable improvements over the basic U-Net architecture with 26.53% in IOU. These results demonstrate that Ens5B-UNet improved the accuracy of MA segmentation and can be integrated into clinical practice. This study introduces a paradigm shift into the application of AI, especially computer vision in eye care for the segmentation of MA.