Diabetic retinopathy (DR) is a leading cause of blindness worldwide. However, 90% of DR caused blindness can be prevented if diagnosed and intervened early. Retinal exudates can be observed at the early stage of DR and can be used as signs for early DR diagnosis. Deep convolutional neural networks (DCNNs) have been applied for exudate detection with promising results. However, there exist two main challenges when applying the DCNN based methods for exudate detection. One is the very limited number of labeled data available from medical experts, and another is the severely imbalanced distribution of data of different classes. First, there are many more images of normal eyes than those of eyes with exudates, particularly for screening datasets. Second, the number of normal pixels (non-exudates) is much greater than the number of abnormal pixels (exudates) in images containing exudates. To tackle the small sample set problem, an ensemble convolutional neural network (MU-net) based on a U-net structure is presented in this paper. To alleviate the imbalance data problem, the conditional generative adversarial network (cGAN) is adopted to generate label-preserving minority class data specifically to implement the data augmentation. The network was trained on one dataset (e_ophtha_EX) and tested on the other three public datasets (DiaReTDB1, HEI-MED and MESSIDOR). CGAN, as a data augmentation method, significantly improves network robustness and generalization properties, achieving F1-scores of 92.79%, 92.46%, 91.27%, and 94.34%, respectively, as measured at the lesion level. While without cGAN, the corresponding F1-scores were 92.66%, 91.41%, 90.72%, and 90.58%, respectively. When measured at the image level, with cGAN we achieved the accuracy of 95.45%, 92.13%, 88.76%, and 89.58%, compared with the values achieved without cGAN of 86.36%, 87.64%, 76.33%, and 86.42%, respectively.

Cataracts are the leading cause of visual impairment worldwide. Examination of the retina through cataracts using a fundus camera is challenging and error-prone due to degraded image quality. We sought to develop an algorithm to dehaze such images to support diagnosis by either ophthalmologists or computer-aided diagnosis systems. Based on the generative adversarial network (GAN) concept, we designed two neural networks: CataractSimGAN and CataractDehazeNet. CataractSimGAN was intended for the synthesis of cataract-like images through unpaired clear retinal images and cataract images. CataractDehazeNet was trained using pairs of synthesized cataract-like images and the corresponding clear images through supervised learning. With two networks trained independently, the number of hyper-parameters was reduced, leading to better performance. We collected 400 retinal images without cataracts and 400 hazy images from cataract patients as the training dataset. Fifty cataract images and the corresponding clear images from the same patients after surgery comprised the test dataset. The clear images after surgery were used for reference to evaluate the performance of our method. CataractDehazeNet was able to enhance the degraded image from cataract patients substantially and to visualize blood vessels and the optic disc, while actively suppressing the artifacts common in application of similar methods. Thus, we developed an algorithm to improve the quality of the retinal images acquired from cataract patients. We achieved high structure similarity and fidelity between processed images and images from the same patients after cataract surgery.

Differentiating arterioles and venules in the fundus image is important for not only various eye diseases but also systemic diseases such as hypertension and ischemic stroke. In this study, we use dual-modal fundus images and develop a cascade refined U-net (CRU-net) to improve the arteriovenous segmentation. Dual-modal fundus images include not only a regular color fundus image (RGB image) but also another two monochromic images acquired using two different wavelengths, 570 nm and 610 nm in our study. The choice of these two wavelengths is based on the absorption spectra of hemoglobin. The two monochromic images provide much richer information on the arteriole and venule. Our proposed CRU-net can fully utilize the information and achieves the state-of-the-art performance on our dual-modal dataset (DualModal2019). The arteriovenous classification accuracy evaluated on the automatically detected vessels is 97.27%, significantly surpassed previous methods. The F1-scores are 77.69% and 79.53% for the arteriole and venule segmentation, respectively. We also test our CRU-net on the public DRIVE dataset with only the color fundus images. We achieve the accuracy of 93.97%, and F1-scores of 73.50% and 75.54% for the arteriole and venule, all of which significantly surpassed previously published methods. Our DualModal2019 dataset with manually annotated arterioles and venules is publicly available.