Drusen are the main aspect of detecting age-related macular degeneration (AMD). Ophthalmologists can evaluate the condition of AMD based on drusen in fundus images. However, in the early stage of AMD, the drusen areas are usually small and vague. This leads to challenges in the drusen segmentation task. Moreover, due to the high-resolution fundus images, it is hard to accurately predict the drusen areas with deep learning models. In this paper, we propose a multi-scale deep learning model for drusen segmentation. By exploiting both local and global information, we can improve the performance, especially in the early stages of AMD cases.

Glaucoma is one of the most common reasons for blindness worldwide, especially in elderly people. Glaucoma can be monitored using visual field (VF) tests. Therefore, predicting the future VF to monitor progression of glaucoma is important. In this paper, we proposed a deep learning model to predict future VF based on previous VF and optical coherence tomography (OCT) images (including thickness map, vertical tomogram, and horizontal tomogram). The image data were analyzed using a ResNet-50 model. Image features and previous VFs were combined, and a long short-term memory (LSTM) network was used to predict future VF. A weighted method was used to detect noisy data. The proposed method was improved when applying weighted loss. The mean absolute error (MAE) was 3.31 ± 1.37, and the root mean square error (RMSE) was 4.58 ± 1.77. The model showed high performance when combining VF data and OCT image data. Furthermore, the model was useful for detecting and re-weighting noisy data.

Diabetic Retinopathy (DR) is one of the major causes of blindness. If the lesions observed in DR occur in the central part of the fundus, it can cause severe vision loss, and we call this symptom Diabetic Macular Edema (DME). All patients with DR potentially have DME since DME can occur in every stage of DR. While synthesizing future fundus images, the task of predicting the progression of the disease state is very challenging since we need a lot of longitudinal data over a long period of time. Even if the longitudinal data are collected, there is a pixel-level difference between the current fundus image and the target future image. It is difficult to train a model based on deep learning for synthesizing future fundus images that considers the lesion change. In this paper, we synthesize future fundus images by considering the progression of the disease with a two-step training approach to overcome these problems. In the first step, we concentrate on synthesizing a realistic fundus image using only a lesion segmentation mask and vessel segmentation mask from a large dataset for a fundus generator. In the second step, we train a lesion probability predictor to create a probability map that contains the occurrence probability information of the lesion. Finally, based on the probability map and current vessel, the pre-trained fundus generator synthesizes a predicted future fundus image. We visually demonstrate not only the capacity of the fundus generator that can control the pathological information but also the prediction of the disease progression on fundus images generated by our framework. Our framework achieves an F1-score of 0.74 for predicting DR severity and 0.91 for predicting DME occurrence. We demonstrate that our framework has a meaningful capability by comparing the scores of each class of DR severity, which are obtained by passing the predicted future image and real future image through an evaluation model.

The performance of existing face age progression or regression methods is often limited by the lack of sufficient data to train the model. To deal with this problem, we introduce a novel framework that exploits synthesized images to improve the performance. A conditional generative adversarial network (GAN) is first developed to generate facial images with targeted ages. The semi-supervised GAN, called SS-FaceGAN, is proposed. This approach considers synthesized images with a target age and the face images from the real data so that age and identity features can be explicitly utilized in the objective function of the network. We analyze the performance of our method over previous studies qualitatively and quantitatively. The experimental results show that the SS-FaceGAN model can produce realistic human faces in terms of both identity preservation and age preservation with the quantitative results of a decent face detection rate of 97% and similarity score of 0.30 on average.