Neovascularization Detection on Optic Disc Region Using Deep Learning.

Diabetic Retinopathy (DR) is one of the biggest eye diseases affecting the worldwide population. The DR presents several damages in the retina depending on its grade of advance, although the damages are asymptomatic in almost all cases. The presence of neovascularization (NV) is considered as the worse stage in the DR, and the patients of this stage require urgent treatment to avoid partial or total blindness. Timely detection of the new vessels in the retina can lead to an adequate treatment to avoid vision loss. In this work, we present automatic detection of neovascularization in the optic disc region (NVD) using a deep learning algorithm. We evaluate several deep neural networks (DNNs) to classify between health optic disc region and NVD. The better DNNs are DenseNet-161 and Efficientnet-B7, which show 93.3% and 92.0% accuracy and 89.5% and 84.2% in sensitivity, respectively. In the computational complexity, DenseNet-161 has a lower number of trainable parameters than that of Efficientnet-B7. To train the DNNs appropriately, we construct a labeled dataset from one of the largest public datasets, bounding NV regions in the retinal images.