Deep Learning-Based Vascular Aging Prediction From Retinal Fundus Images

Abstract

Purpose: The purpose of this study was to establish and validate a deep learning model to screen vascular aging using retinal fundus images. Although vascular aging is considered a novel cardiovascular risk factor, the assessment methods are currently limited and often only available in developed regions.

Methods: We used 8865 retinal fundus images and clinical parameters of 4376 patients from two independent datasets for training a deep learning algorithm. The gold standard for vascular aging was defined as a pulse wave velocity ≥1400 cm/s. The probability of the presence of vascular aging was defined as deep learning retinal vascular aging score, the Reti-aging score. We compared the performance of the deep learning model and clinical parameters by calculating the area under the receiver operating characteristics curve (AUC). We recruited clinical specialists, including ophthalmologists and geriatricians, to assess vascular aging in patients using retinal fundus images, aiming to compare the diagnostic performance between deep learning models and clinical specialists. Finally, the potential of Reti-aging score for identifying new-onset hypertension (NH) and new-onset carotid artery plaque (NCP) in the subsequent three years was examined.

Results: The Reti-aging score model achieved an AUC of 0.826 (95% confidence interval [CI] = 0.793–0.855) and 0.779 (95% CI = 0.765–0.794) in the internal and external dataset. It showed better performance in predicting vascular aging compared with the prediction with clinical parameters. The average accuracy of ophthalmologists (66.3%) was lower than that of the Reti-aging score model, whereas geriatricians were unable to make predictions based on retinal fundus images. The Reti-aging score was associated with the risk of NH and NCP (P < 0.05).

Conclusions: The Reti-aging score model might serve as a novel method to predict vascular aging through analysis of retinal fundus images. Reti-aging score provides a novel indicator to predict new-onset cardiovascular diseases.

Translational Relevance: Given the robust performance of our model, it provides a new and reliable method for screening vascular aging, especially in undeveloped areas.

Introduction

Vascular aging may be a biomarker of age-related health impairment and considered a novel cardiovascular risk factor, characterized by arterial stiffness.¹ Pulse wave velocity (PWV) is a noninvasive measurement of arterial stiffness because it is closely related to the mechanical characteristics of the arterial wall.^2–4 However, the equipment required for PWV measurement can be costly and may not be readily available in all healthcare facilities, especially in resource-limited regions. These limitations make it challenging for PWV to become a screening method for vascular aging and cardiovascular disease in large-scale populations. 

The eyes are highly vascularized photosensitive organs. Although the volume of both eyes is estimated to be approximately 0.2% to 0.3% of the total body, the blood flow to the eyes constitutes a relatively rich proportion of the cardiac output, approximately 4% to 5%.⁵ The eyes allow noninvasive, direct observation of blood vessels, offering a unique insight into vascular dynamics.^6–8 Retinal microvascular abnormalities have been demonstrated to be associated with coronary heart disease and stroke.^9,10 Central retinal venular equivalent is associated with the incidence of heart failure, and wider central retinal venular equivalent and narrower central retinal arteriolar equivalent are correlated with left ventricular size, left ventricular hypertrophy, and diastolic-systolic dysfunction.¹¹ These changes in the retinal microvasculature are indicative of compromised blood flow and vascular integrity, which could be important indicators of vascular aging.¹² The retinal manifestations can be easily visualized using retinal fundus images, which is widely available across various healthcare institutions. 

Deep learning operates by using multiple layers of computation to enable the algorithm to autonomously extract predictive features from samples, eliminating the need for manual feature engineering.¹³ Deep learning models, especially convolutional neural networks, have demonstrated remarkable capabilities in detecting and classifying various eye conditions, such as diabetic retinopathy,¹⁴ age-related macular degeneration,¹⁵ and glaucoma,¹⁶ as well as systemic diseases like cardiovascular diseases and hypertension,^17,18 through distinctive retinal biomarkers. “Retinal age gap,” determined as retinal age derived from fundus images minus chronological age, was recently preserved as a robust biomarker linked to the aging process. Recent evidence suggests that the retinal age gap may be associated with Parkinson's disease, renal failure, stroke, and mortality, indicating that the retina contains signaling of aging.^19–21 However, the role of deep learning algorithms in retinal fundus images for vascular aging remains unexplored. 

We used retinal fundus images to train, validate, and test a model, defining the probability of the presence of vascular aging as deep learning retinal vascular aging score, namely Reti-aging score. Regression models were subsequently constructed by integrating clinical parameters with Reti-aging score. Furthermore, we analyzed the correlation between the Reti-aging score and various cardiovascular risk factors, as well as its predictive capacity for new-onset cardiovascular diseases. 

Methods

Study Design and Population

In this retrospective, bicenter study, we collected 8865 retinal fundus images and clinical parameters from 4376 patients to train, validate, and test a deep learning model for predicting vascular aging. Retinal fundus images from both eyes were used for analysis, and a small subset of images was from the same eye at the same time points. Data from two independent health-screening centers were used, including the main department of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (Dataset 1: developmental and internal test data) and Optics Valley department of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (Dataset 2: external test dataset). The patient population and collection process of these two datasets were independent of each other. 

For all participants, we collected retinal fundus images centered at the midpoint between the macula and optic disc. Physiological parameters, including age, gender, body mass index (BMI), systolic blood pressure, diastolic blood pressure, and PWV, were also collected at the same time. PWV was collected from all participants according to the recommended method.² We simultaneously measured PWV on both the right and left sides, and the average values in each individual were used for statistical analysis. Other clinical information, including hypertension, diabetes, current history of smoking, and alcohol consumption, were obtained through patient-reported recall. 

The retinal fundus images were obtained by using standard photographic and darkroom procedures, using a Canon CR-2 digital retinal camera (Canon Inc., Tokyo, Japan) in the dataset 1 and Canon CR-2 AF digital retinal camera (Canon Inc.) in dataset 2. PWV was measured using the Vascular Profiler BP-203RPEIII system (Omron Healthcare, Kyoto, Japan) on patients using the approach described previously,³ as shown in Figure 1A. To maximize the sample size of the developmental dataset, all participants older than 18 with matched retinal fundus images and PWV measurements were included in the study. The exclusion criteria in this study were defined as follows: (1) poor quality of images: unclear optic disc or major retinal vessels; (2) any missing data in clinical parameters: including age, gender, systolic and diastolic blood pressure, BMI, hypertension, diabetes, current smoker, and alcohol consumption; (3) those suffer from severe corneal edema, cataract, vitreous hemorrhage, and other ophthalmic diseases that prevent observation of the retina; (4) those suffer from serious systemic diseases or malignant tumors. 

Figure 1.

 

Overview of the deep learning models development. (A) Schematic representation of the PWV examination method. (B) Image labeling. Retinal fundus images were categorized and labeled based on patient PWV measurements. (C) Model construction. A deep learning model was constructed using retinal fundus images. The predictive ability of Reti-aging score for vascular aging was compared with clinical parameters using Logistic Regression.

This study was approved by the medical ethics committee of Tongji hospital and followed tenets of the Declaration of Helsinki. This study was registered at Chinese Clinical Trial Registry (ChiCTR2100054870). 

Definitions in This Study

According to guidelines, vascular aging was defined as PWV ≥1400 cm/s.^2,22,23 The Reti-aging score was defined based on a probability score derived from our deep-learning algorithm of absence or presence of vascular aging. In this study, new-onset diseases were defined as the presence of diseases within the subsequent three-year follow-up. 

Deep Learning Algorithm for Predicting Vascular Aging

In this study, we aimed to train a deep learning model for analyzing retinal fundus images to predict vascular aging. Dataset 1 was used for training and validation, and dataset 2 was used for testing. The retinal fundus images were labeled into binary classification based on the presence or absence of vascular aging (Fig. 1B). 

Our framework included two models, one was deep learning model, and the other was regression model which was built on the combination of clinical information and the output probability of the deep learning model. Specifically, our deep learning model included five sections. For the first four sections, each section consisted of three layers as follows. The first layer was a convolutional layer that convolved the output of the previous layer with a set of kernels to extract a set of features at different locations of the original image and obtain the representation which was called the feature maps. The second layer was the batch normalization layer, which normalized the values of different feature maps in the previous layer, followed by an activation function called rectified linear unit (RELU). The third layer was the max pooling layer, which lowered the computational burden by compressing the feature map processed by previous layers. The last section consisted of four layers including a convolutional layer, a batch normalization layer with RELU, a global average pooling layer, and a softmax layer. The global average pooling can largely decrease the number of parameters to reduce the risk of overfitting. The softmax layer provided the posterior probability of each class by a softmax function. The numbers in the picture like “256 × 256 × 3” represented the length, width, and channels of the feature maps, respectively, and N referred to the number of classes of the input images. The schematic diagram of the model architecture is depicted in Figure 1C. 

Our Reti-aging score model was implemented based on PyTorch and run on two NVIDIA RTX 2080Ti graphics cards. Our logistic regression model was implemented based on R. In our proposed Reti-aging score model structure, the kernel size of each convolution was fixed to 3, with the convolution stride fixed to 1, and the pooling size of max-pooling layer was fixed to 2, with the pooling stride fixed to 2. An effective optimization method named Adam was adopted to achieve efficient computing. A weighted loss function named as focal loss was adopted during training to achieve better performance by focusing on the samples that were not easily classified. The hyperparameters of our proposed Reti-aging score model were set to {8, 0.00005, 200}, which denoted the batch size, learning rate, and training epoch, respectively. 

Performance Evaluation of the Reti-Aging Score Model

The ability of our model to predict vascular aging was compared with that of models on other cardiovascular disease risk factors. Additionally, to enhance the interpretability, saliency maps highlighting pivotal regions in the final image classification were generated using CAM methodology. The six specialists, comprising three ophthalmologists and three geriatricians, were tasked with diagnosing vascular aging based on retinal fundus images for evaluating the performance. 

Correlation Analysis Between the Reti-Aging Score and New-Onset Cardiovascular Diseases

Dataset 1 is sourced from a health examination dataset spanning from 2013 to 2020, wherein some individuals underwent regular follow-up every year. There were 733 individuals with records of more than three years, encompassing 7215 fundus images from 2015 samples. Hypertension diagnosis is established through either patient self-reporting of hypertension history or use of hypertension-lowering medications. Carotid artery plaques were assessed using carotid ultrasound scanning. New-onset hypertension (NH) and new-onset carotid artery plaque (NCP) were identified if they were present at least once in the subsequent three years of follow-up. 

To assess whether the Reti-aging score is a potential risk factor for evaluating cardiovascular diseases, we performed an analysis of various cardiovascular risk factors within the stratified three-classification framework of Reti-aging score. Furthermore, we investigated the relationship between Reti-aging score and the incidence of NH and NCP using logistic regression analysis. 

Statistical Analysis

We reported the mean (± standard deviation [SD]) for description of distributions. Categorical variables were described using count (percentage). Continuous data were tested for normality using a Shapiro–Wilk normality test. Subsequently, one-way analysis of variance or Kruskal-Wallis test was used for normally or non-normally distributed data, respectively. Fisher's exact test was used for categorical variables. Independent sample t-test or Mann-Whitney U test was conducted for normally or non-normally distributed data, respectively. Statistical significance was determined at a P value < 0.05. Data analysis was performed using IBM SPSS Statistics version 26 (SPSS Inc., Chicago, IL, USA). Intereye correlation was not accounted for retinal images from both eyes might contain varied information. Area under the curve (AUC), Net Reclassification Index (NRI), sensitivity, specificity, and F1 were used to evaluating the model's performance. NRI represents the extent to which the new model improves on a baseline model. It is calculated as the difference between the sum of sensitivity and specificity of a diagnostic test and another diagnostic test. Logistic regression analyses were presented as odds ratio with 95% confidence intervals (95% CI). 

Results

Basic Characteristics of the Developmental Dataset

In this study, dataset 1 is a cross-sectional and longitudinal dataset. We collected 31,045 retinal fundus images, from 26,085 patients, taken between 2013 and 2020. A total of 2772 samples did PWV measurements, paired with 4979 images. After applying the exclusion criteria screening, a dataset comprising 4874 retinal fundus images from 2717 patients was used for the development and internal validation of the model, with an average follow-up duration of 3.9 years. The flowchart is shown in Figure 2. 

Figure 2.

 

Study profile of dataset selection.

In dataset 1, the mean age of subjects was 48.4 (±9.7), and the mean PWV value was 1344.5 (±251.8), with 940 (34.5%) subjects exhibiting vascular aging. Other basic characteristics of the datasets are shown in Table 1. We depicted the distributions of PWV measurements and age for both male and female groups. The scatter plot revealed that the PWV was higher in the male group compared to the female group in our dataset (Fig. 3). 

Table 1.

 

Baseline Characteristics of Datasets

 

Figure 3.

 

Distributions of PWV and age in male and female groups. Among individuals of the same age, males exhibit a slightly higher PWV compared to females.

The subjects were randomly divided into two sets: the training set and the validation set at the ratio of 8:2. Images captured from the same subject were grouped together within a dataset to ensure their exclusion from both the training and testing groups, thereby preventing any overlap in the utilization of these images within the model. The probability score derived from our deep-learning algorithm of absence or presence of vascular aging was recorded as Reti-aging score. 

Basic Characteristics of the External Test Dataset

To further test generalizability of this model, we used another population-based dataset 2 for external test. A total of 7135 retinal fundus images were obtained from 2871 patients. After the application of matching and exclusion criteria, we established an external test dataset, encompassing 3991 images derived from a cohort of 1659 subjects. The flowchart is shown in Figure 2. 

In dataset 2, the mean age of subjects was 47.1 (±10.6), with 836 (50.4%) subjects exhibiting vascular aging. Other basic characteristics of the datasets are shown in Table 1. 

Performance of the Reti-Aging Score Model in Predicting Vascular Aging

The outcomes of Reti-aging score model and other clinical parameters in predicting vascular aging are illustrated in Table 2. The receiver operating characteristic (ROC) curves in the internal and external test set are displayed in Figure 4. 

Table 2.

 

AUC of Reti-Aging Score Versus Clinical Parameters in Predicting Vascular Aging

 

Figure 4.

 

ROC curves of Reti-aging score model and clinical parameter models. (A) ROC curves for internal test dataset. (B) ROC curves for external test dataset.

In the internal test dataset, the Reti-aging score model achieved better performance (0.826 [0.793–0.855]), compared with single-parameter models, such as gender (0.61 [0.57–0.647]), age (0.795 [0.752–0.835]), hypertension (0.635 [0.592–0.677]), diabetes (0.55 [0.524–0.578]), BMI (0.665 [0.617–0.716]), systolic blood pressure (0.803 [0.758–0.843]), diastolic blood pressure (0.779 [0.734–0.818]). When Reti-aging score was added to all clinical parameters model, the model's AUC increased from 0.874 (95% CI = 0.841–0.903) to 0.882 (95% CI = 0.861–0.904), with an NRI of 0.179 (0.104–0.25), P < 0.0001. 

In the external test dataset, the Reti-aging score model achieved an AUC of 0.779 (0.765–0.794), compared with single-parameter models, such as gender (0.487 [0.466–0.509]), age (0.715 [0.689–0.739]), hypertension (0.574 [0.558–0.59]), diabetes (0.545 [0.533–0.557]), BMI (0.549 [0.522–0.577]), systolic blood pressure (0.762 [0.74–0.785]), diastolic blood pressure (0.74 [0.717–0.763]). When Reti-aging score was added to all clinical parameters model, the model's AUC increased from 0.799 (95% CI = 0.779–0.82) to 0.825 (95% CI = 0.812–0.837), with an NRI of 0.242 (0.214–0.271), P < 0.0001. These results suggest that the Reti-aging score might play an independent role in prediction of vascular aging. 

A detailed comparison of the predictive capabilities of the Reti-aging score model and all clinical parameter models in predicting vascular aging, such as accuracy, sensitivity, specific, and F1 are shown in Table 3. 

Table 3.

 

Performances of Reti-Aging Score Versus Multi-Parameters Model in Predicting Vascular Aging

 

In the human-deep learning model comparison, we randomly selected 80 retinal fundus images and asked ophthalmologists and geriatricians to independently assess vascular aging based on the images. The average accuracy of ophthalmologists was 66.3%, lower than the performance of Reti-aging score model, whereas the geriatrists could not make the prediction on the basic of retinal fundus images. 

Visualization of Evidence for Vascular Aging Prediction

To enhance the explicability of the Reti-aging score model and elucidate its diagnostic mechanism, integrated gradients were used to produce saliency maps. Representative examples of primary retinal fundus images along with their associated saliency maps are exhibited in Figure 5. 

Figure 5.

 

Saliency maps of the Reti-aging score model. (A) Examples of original retinal fundus images. (B) Examples of saliency maps. (a) Normal sample. (b) Vascular aging sample.

The knowledge obtained from saliency maps indicates that the Reti-aging score model was focused on distinct regions in cases of vascular aging and nonvascular aging. In the assessment of non-vascular aging, the areas where retinal arteries and veins are located received attention. This region holds significance for ophthalmologists in diagnosing retinal pathologies, suggesting the Reti-aging score model's emphasis on clinically relevant features. Interestingly, under the conditions of vascular aging, attention was directed toward regions outside the area of retinal arteries and veins. The unexpected attention indicates the potential of the Reti-aging score model to capture subtle retinal variations under the influence of vascular aging, a phenomenon typically difficult for ophthalmologists to directly observe. 

Correlation of Reti-Aging Score With New-Onset Cardiovascular Diseases

We conducted an analysis of cardiovascular risk factors within the stratified framework of three classification of Reti-aging score. The outcomes of this investigation are detailed in Table 4. The results demonstrated a significant increasing trend in the mean values of some clinical parameters across the tripartite divisions. For instance, within the cohort of three patient groups, the mean PWV values were 1150.1 ± 159.6, 1353.4 ± 239.0, and 1680.4 ± 324.5 (P < 0.05); mean ages were 35.6 ± 7.7, 49.4 ± 8.3, and 61.7 ± 8.0 (P < 0.05); mean systolic blood pressures were 114.9 ±12.5, 123.7 ± 15.2, and 134.8 ± 14.7 (P < 0.05); and mean diastolic blood pressures were 74.0 ± 9.9, 81.2 ± 11.3, and 85.5 ± 10.8 (P < 0.05). Notably, we observed an elevated trend in the prevalence of cardiovascular and metabolic diseases. For instance, within the three patient groups, prevalence of hypertension was 2.2%, 21.7%, and 42.0% (P < 0.05); prevalence of carotid artery plaque was 1.7%, 12.2%, and 23.3% (P < 0.05); prevalence of diabetes was 2.0%, 5.9%, and 24.0% (P < 0.05); and prevalence of hyperlipidemia was 23.1%, 32.6%, and 38.7% (P < 0.05). Moreover, we noted an elevated proportion of male patients, consistent with the observed phenomenon of higher male PWV compared to female PWV at equivalent ages (Fig. 3). As the Reti-aging score increases, the prevalence of smoking and alcohol consumption among patients also rises. 

Table 4.

 

Characteristics of Clinical Information in Reti-Aging Score Stratification

 

The mean values of Reti-aging score for the NH group and non-NH group were 0.52 ± 0.10 and 0.45 ± 0.11, respectively. Additionally, the mean values of Reti-aging score for the NCP group and non-NCP group were 0.51 ± 0.10 and 0.45 ± 0.11, respectively. The results are shown in Figure 6. We evaluated the predictive capability of Reti-aging score for hypertension and carotid artery plaque, with AUCs of 0.703 (95% CI = 0.678–0.720) and 0.705 (95% CI = 0.684–0.726), respectively. 

Figure 6.

 

Performance of Reti-aging score predicting new-onset cardiovascular diseases. (A) Bar graphs for the NH group and none NH group. (B) Schematic representation of difference in NH group and none NH group. (C) Schematic representation of difference in NCP group and none NCP group. (D) Bar graphs for the NCP group and none NCP group.

To evaluate whether Reti-aging score serves as a potential cardiovascular risk factor, we further used logistic regression analysis with adjustment for age, gender, BMI, systolic pressure and diastolic pressure to examine its association with NH and NCP. Our findings revealed that the odds ratio was 3.059 (95% CI = 0.362–25.852, P < 0.05) for Reti-aging score and NH, and 1.255 (95% CI = 0.347–4.529, P < 0.05) for Reti-aging score and NCP. 

Discussion

In this study, we developed a deep-learning algorithm to predict vascular aging through retinal fundus images. In contrast to individual clinical parameters like age, gender, hypertension, diabetes, BMI, or blood pressure, the Reti-aging score exhibited better predictive capability. When Reti-aging score was added to all clinical parameters model, the model's predictive capacity was elevated, suggesting that the Reti-aging score has predictive capability independent of previously reported parameters. On analysis, we have noted an upward trend in multiple cardiovascular risk factors under the three-classification of Reti-aging score. Our investigations revealed a noteworthy association between Reti-aging score and the occurrence of NH and NCP in patients. This outcome revealed that Reti-aging score not only serves as a diagnostic tool for vascular aging but also demonstrates potential for forecasting the new-onset cardiovascular diseases. 

The novelty of this study lies in the initial development and validation of a deep learning algorithm for prediction vascular aging based on retinal fundus images, using a bicenter, population-based dataset. Furthermore, this study confirmed the Reti-aging score's potential to predict the emergence of new-onset cardiovascular diseases. A recent study introduced a deep learning model that used ultra-widefield pseudo-color retinal images to predict PWV. However, this study was constrained by a restricted sample size, a lack of external validation, and the investigation into the model's capability to predict new-onset cardiovascular diseases.²⁴ 

We have determined that retina images offer more information and predictive signals for vascular aging than some individual clinical parameter, such as age or hypertension (Table 2). Abundant prior evidence has indicated that age and hypertension are pivotal risk factors for vascular aging.^25,26 However, our model outcomes suggest that the retina, a vasculature-rich neural peripheral tissue, might encompass various information including age, blood pressure, BMI, and blood lipids (Table 4). These finding are consistent with earlier research proposing that retinal images containing cardiovascular risk information.^9,19 Several deep learning models have substantiated that the algorithms can capture specific features within the retina, such as vascular/macular region and optic disc, enabling the evaluation of cardiovascular disease indictors and cardiovascular events.^17,27,28 

PWV reflects the distance over which a pulse wave propagates along blood vessels within a unit of time.³ The elevation in PWV, often linked to vascular aging, is viewed as a manifestation of vascular functional alteration.²² Some studies posit that this functional alteration occurs before structural vascular wall lesions, potentially instigating endothelial injury and the formation of atherosclerotic plaques.^2,4 The predictability of the Reti-aging score for the new onset of cardiovascular diseases is consistent with existing this evidence. 

The retina has two sets of blood vessels, including the choroidal vessels that supply the outer retina and the retinal vessels that supply the inner retina.²⁹ The retinal vessels can be divided into superficial vessels, intermediate capillary layer, and deep capillary layer.³⁰ In the condition of nonvascular aging, the assessment of retinal large vessels plays a predictive role in determining outcomes, owing to the fact that retinal vasculature reflect the systemic vascular status.³¹ Conversely, under conditions of vascular aging, attention shifts beyond the retinal large vessels. The widespread distribution of capillaries in the non-large vessel regions of the retina may explain this phenomenon. Vascular aging affects both large blood vessels and microvessels.³² During the progression of vascular aging, microvessels might be influenced earlier than those larger vessels. Retinal microcirculation abnormalities characterized by changes in retinal color may explain. 

Although it may be limited medical resources, the Reti-aging score model holds the potential to assist clinicians in conducting vascular aging screening and risk analyses for new-onset cardiovascular diseases. This offers a novel approach to addressing vascular aging screening in rural areas. In the future, through algorithm refinement, the prospect of attaining a multifaceted evaluation of systemic diseases becomes feasible. 

There are several limitations in this study. First, the two databases used in this study both originated from Wuhan, Hubei Province, China, which may introduce geographical and population bias. Second, causal relationships between retinal features and vascular aging were not explored. Third, potential ocular features associated with vascular aging were not further validated through other ophthalmic examinations such as optical coherence tomography and ocular angiography. 

Conclusions

We developed a deep learning algorithm to predict vascular aging using retinal fundus images. Reti-aging score exhibited better predictive capabilities for vascular aging in comparison to some clinical parameters alone. Building on this discovery, we extended our predictions to encompass new-onset cardiovascular diseases within the subsequent three years using the Reti-aging score. Consequently, the Reti-aging score model holds the potential to serve as an innovative imaging modality for both vascular aging screening and prognosticating cardiovascular diseases. 

Acknowledgments

Supported by the National Natural Science Foundation of China under Grant (81974136), National Key R&D Program of China (2020YFC2008000), Key R&D Plan Projects in Hubei Province (2022BCA011). 

Disclosure: R. Wang, None; Y. Tan, None; Z. Zhong, None; S. Rao, None; Z. Zhou, None; L. Zhang, None; C. Zhang, None; W. Chen, None; L. Ruan, None; X. Sun, None