Diabetic retinopathy (DR) causes blindness in young adults worldwide. Thanks to early detection, patients with diabetic retinopathy can be properly treated in time, and the deterioration of diabetic retinopathy can be prevented. Early detection is therefore essential for screening the DR disease. To perform a supervised DR classification, we need labeled images, otherwise we are required to conduct a traditional manual diagnosis, with the help of an ophthalmologist, which is timeconsuming and costly expensive. The question here is what we can do if we have unlabeled images? We can benefit from the knowledge of a pre-trained classifier using a labeled base, but the problem is that the images in the datasets may come from different domains. To solve all these problems, we propose a three-step method: first, train a classifier using a labeled dataset; second, adapt the unlabeled dataset (source domain) to the labeled dataset (target domain) using Least Squares Cycle-GAN; and finally, classify these adapted images using the pretrained classifier on the target domain. Following this procedure, the results show that we succeeded in classifying 77% of unlabeled images, meaning that the ophthalmologist can concentrate on the remaining 23% of cases for diagnosis, which is clear time saving. This approach is contribution to the existing supervised learning and transfer learning methods that are requiring a lot of labeled data, which is always not available in DR screening due to the time consuming of image fundus annotating.