Retinal Age Recognition Based On Deep Learning Models Using Fundus Color Photographs

As the population aging becomes a social transformation on a global scale,age-related diseases are becoming more and more common,posing a significant threat to human health.The retina,which is the only tissue in the body that allows visualization of blood vessels,provides a unique "window" for the clinical evaluation of the risk of systemic vascular and neurological diseases associated with aging.Previous studies have shown that people whose retinal aging differs significantly from their actual age are 49% to 67% more likely to die due to any disease besides cardiovascular disease and cancers.The association between retinal aging and death from other diseases suggests that for every additional year of difference between retinal age and actual age,the risk of premature death increases by 2% to 3%.Furthermore,previous research has shown a strong correlation between the difference in retinal age and physiological age and the risk of metabolic syndrome,inflammation,kidney failure,arterial stiffness,and cardiovascular disease incidence.Therefore,identifying retinal age has significant values for both academic research and medical application.The purpose of this paper is to design and implement an efficient retina age classification network that can fully explore the relationship between retinal aging and date-of-birth age,evaluate physical health status,and provide specific innovative points and work as follows:(1)Classic convolutional residual networks are used for age recognition.Retina age recognition was performed on a preprocessed population aged 26-65 using four different network architectures.These four neural network architectures are: a single Res Net convolutional neural network,a Res Net convolutional neural network with combined ECA attention mechanism and SE attention mechanism,an SE-Inception-V4 neural network,and a Vi T neural network with self-attention mechanism.The experiment was carried out with an eight-category training according to five-year intervals.The results showed that there is a certain correlation between retinal aging and age,but the above classic network classification methods have low recognition accuracy and poor robustness due to the lack of feature information related to retinal age in fundus images.(2)Attention mechanisms are used to visualize and locate retinal age features.To address the problems in classic networks,it is necessary to identify the feature information related to retinal age in fundus images to guide neural network age recognition training and improve its training efficiency.Therefore,this paper proposes a method of low-dimensional classification localization of retinal age features,which divides the fundus image data into large categories according to ten-year age intervals,and then uses a residual network based on channel attention for four-category training to visually locate the retinal age-related regional features.The results showed that the macular region and the area near the fundus blood vessels are relatively rich in age-related information.(3)Design and use a dual-tower model for feature fusion training.In order to fully utilize the rich feature information of the macular region and retinal vascular region mentioned above and improve the accuracy of neural network retinal age recognition,this paper proposes a feature extraction method for the retinal vessels and their adjacent areas combining masking and image morphology techniques to extract the age features of the retinal vessels and their adjacent areas,which are then used as the main features along with the optic disc and macular age features.The features are input into the dual-tower neural network model for feature fusion training.First,the retinal vascular area is segmented,dilated and closed using image morphology techniques to obtain vessel and adjacent region images.Then,the feature images are input into two main networks for feature extraction,followed by feature fusion in the feature fusion unit.Finally,the training result is obtained through an eight-category classifier.The results showed that the dual-tower neural network model based on feature fusion and attention mechanisms further improved the recognition accuracy of the eight-category retina age classification.The dual-tower neural network model proposed in this paper was tested on a test set containing more than 1600 fundus images,achieving a recognition accuracy of 86.28%,with a statistically significant improvement of2%～5%in recognition accuracy,enabling more precise identification of the subject’s retinal age and providing a more effective tool for early warning and health management of age-related chronic diseases.(4)Use adaptive threshold segmentation and morphological methods for pupil localization and tracking.Based on the combination of adaptive threshold segmentation and morphological methods in Open CV,an interactive software is designed for pupil tracking by automatically tracking and calculating the pupil position.This method is convenient for the operator to collect fundus images efficiently and accurately,decreasing the risk of multiple alignment correction of eye position and re-acquisition of fundus images due to poor pupil tracking,thus improving the efficiency of fundus image acquisition.