Using Fully Convolutional Networks For Semantic Segmentation of Diabetic Retinopathy Lesions in Retinal Images

In this study, we investigate the use of a fully convolutional network (FCN) architecture for semantic segmentation of specific lesions in retinal fundus images. Using data obtained from publicly available databases, pre-trained FCN-8s networks were fine-tuned to detect micro-aneurysms (MA), hemorrhages (HEM), hard exudates (HE) and soft exudates (SE). Performance was evaluated with a standard F1-score as well as a hit and miss metric on ground truth test data for two networks trained on different versions of the original dataset. The preliminary results presented in this paper are encouraging with regards to the use of deep-learning algorithms for automatic detection and segmentation of these specific retinal lesions, suggesting that this method should be explored in the future for use in automatic detection and segmentation of retinal lesions in diseases such as diabetic retinopathy (DR). At test time, the best performing network achieved a sensitivity of 0.80, 0.85, 0.70 and 0.81 respectively for detection of each lesions type: MA, HEM, HE and SE, according to the hit and miss metric and a recall of 0.59 for red lesion (MA and HEM) for pixel-level segmentation.