Early diagnosis through regular screening is important for preventing Diabetic retinopathy (DR), but it is time consuming for ophthalmologists to diagnose efficiently. 

Previous studies from the Shenzhen Institutes of Advanced Technology (SIAT) of the Chinese Academy of Sciences has found that computer-aided diagnosis system using deep-learning technology could realized high-level features learning from DR images, which could automatically analyzed the possibility of DR.

However, deep learning models are usually trained on a large amount of labeled DR data. Moreover, for the labelling process of DR images, the grading of DR requires the clinician to extract the lesions and measure the area of the lesions manually, which is highly time consuming. Due to the lack of high-quality labeled data in real applications, it is difficult to apply the general deep learning method (such as GoogleLeNet, ResNet) for DR diagnosis. 

Their current research solved these difficulties. 

In this study, the scientists proposed a multi-channel based generative adversarial network (MGAN) with semi-supervision to grade DR, which could make full use of labeled data and unlabeled data to recognize DR automatically without losing the original DR features. 

To deal with the challenge that the effective DR features (e.g., exudates, microaneurysms and bleeding points) are diffuse in the high-resolution fundus images, the researchers developed a multi-channel based GAN model, which could generated a series of sub-fundus images included effective local features. All the sub-fundus images were then combined to obtain the most representative features of the entire fundus image.  

Besides, the researchers incorporated a feature extraction scheme into the proposed multi-channel based GAN framework. This scheme reduced the noise from the original fundus images and extracted the scattering lesion features, which improved the performance of discriminator. 

In order to demonstrate the advantage of the proposed method, the researchers employed 100 labeled DR samples but a large number of unlabeled DR samples to train the model.  

The results exhibited that proposed model could deal with a classification problem when the labeled samples were limited, and outperformed the other representative models in terms of accuracy(96.6%), area under the curve (AUC, 98.3%), sensitivity(97.0%) and specificity(96.8%).  

The study was published in IEEE Transactions on Automation Science and Engineering.