Association of Retinal Biomarkers With the Subtypes of Ischemic Stroke and an Automated Classification Model

Abstract

Purpose: Retinal microvascular changes are associated with ischemic stroke, and optical coherence tomography angiography (OCTA) is a potential tool to reveal the retinal microvasculature. We investigated the feasibility of using the OCTA image to automatically identify ischemic stroke and its subtypes (i.e. lacunar and non–lacunar stroke), and exploited the association of retinal biomarkers with the subtypes of ischemic stroke.

Methods: Two cohorts were included in this study and a total of 1730 eyes from 865 participants were studied. A deep learning model was developed to discriminate the subjects with ischemic stroke from healthy controls and to distinguish the subtypes of ischemic stroke. We also extracted geometric parameters of the retinal microvasculature at different retinal layers to investigate the correlations.

Results: Superficial vascular plexus (SVP) yielded the highest areas under the receiver operating characteristic curve (AUCs) of 0.922 and 0.871 for the ischemic stroke detection and stroke subtypes classification, respectively. For external data validation, our model achieved an AUC of 0.822 and 0.766 for the ischemic stroke detection and stroke subtypes classification, respectively. When parameterizing the OCTA images, we showed individuals with ischemic strokes had increased SVP tortuosity (B = 0.085, 95% confidence interval [CI] = 0.005–0.166, P = 0.038) and reduced FAZ circularity (B = −0.212, 95% CI = −0.42 to −0.005, P = 0.045); non–lacunar stroke had reduced SVP FAZ circularity (P = 0.027) compared to lacunar stroke.

Conclusions: Our study demonstrates the applicability of artificial intelligence (AI)–enhanced OCTA image analysis for ischemic stroke detection and its subtypes classification. Biomarkers from retinal OCTA images can provide useful information for clinical decision–making and diagnosis of ischemic stroke and its subtypes.

The medical and social burden of stroke has increased rapidly over the last 2 decades,¹ and recent reports have shown that stroke is now the leading cause of mortality and disability worldwide.² Ischemic stroke accounts for about 80% of all strokes. Ischemic stroke may further be classified into 2 subtypes: lacunar stroke and non–lacunar stroke, with lacunar stroke accounting for up to 25% of all ischemic stroke cases and non–lacunar stroke accounting for 75%.³ Accurately distinguishing these two subtypes is pivotal due to their different treatment requirements. Furthermore, the timely diagnosis is also a crucial factor influencing the outcomes of patients. In current practice, diagnosis of ischemic stroke has been facilitated by computed tomography (CT) and magnetic resonance imaging (MRI). These imaging modalities have demonstrated great capability at detecting cerebrovascular changes and diagnosing stroke, but these tools are also suffering from low imaging resolution, are time–consuming, at a high cost, and are sometimes invasive. 

The neurosensory retina is a developmental outgrowth of the brain. The retina and the brain share many features, such as precise neuronal cell layers, blood barriers, embryologic origin, and microvasculature.⁴ Unlike the other tissues of the central nervous system, the retinal blood vessels can be visualized directly using ophthalmic imaging tools. For example, several studies^5,6 analyzed the color fundus images and showed individuals with stroke had reduced retinal vessel fractal dimension and increased vessel tortuosity when compared with controls. It is noted though, that color fundus imaging modality cannot capture the retinal capillaries in sufficient detail, which implies it is hard to capture the microvascular network. Recently, several studies utilized optical coherence tomography angiography (OCTA) to investigate the associations between microvasculature changes and stroke, as it enables the visualization of retinal microvasculature in different retinal layers with high resolution. Wiseman et al.⁷ reported that OCTA–derived metrics provide evidence of retinal microvascular abnormalities that may underlie cerebral small vessel disease (SVD) lesions in the brain, and SVD is a major cause of stroke. Liang et al.⁸ reported that individuals with stroke have reduced retinal microvascular densities compared with controls. Zhang et al.⁹ reported that the deep retinal capillary density in two –stroke subgroups was reduced in all regions except the inferior region, and the fractal dimension in specific subregions of the deep retina was lower in the non–lacunar group. Duan et al.¹⁰ demonstrated that optical coherence tomography (OCT) showed specific damage patterns in retinal microvascular and macular morphology in different subtypes of ischemic stroke. Despite the usefulness that OCTA demonstrated, almost all these previous studies used built–in software from OCTA devices, which provided limited information on the retina, and sometimes resulted in inconsistent results. Overall, it is suggested that OCTA imagery has the potential to be a tool to detect microvascular changes in patients with ischemic stroke. 

Advances in machine learning have spawned significant efforts to incorporate machine learning with the clinical diagnosis of stroke in recent years, and published works have largely focused on neuroimaging such as MRI and CT.^11–14 In this study, we introduced a novel deep–learning model to identify ischemic strokes and classify them into lacunar and non–lacunar strokes using OCTA imagery. In addition, we used a well–established segmentation method to detect retinal vessels and obtained two vessel–related parameters that characterize both the retinal microvasculature and the foveal avascular zone. We further explored the associations of retinal microvascular changes in different study groups and the results show consistency with the conclusions of the interpretation of our deep learning model. 

Materials and Methods

Study Participants

This study includes two cohorts, one for ischemic stroke identification and another for lacunar and non–lacunar stroke classification. Participants in our study were enrolled in a cross–sectional study design approved by the Ethics Committee of the Institute of Biomedical Engineering, Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, and West China Hospital of Sichuan University Ethics Committee (Ethics number 2020[922]). This study followed the tenets of the Declaration of Helsinki. Written informed consent was obtained from each participant before enrolling. 

To build the ischemic stroke identification cohort, patients from the Department of Neurology, West China Hospital of Sichuan University, with first–ever stroke within 21 days of onset were recruited from December 2020 to December 2022. The inclusion criteria are as follows: (1) age ≥ 18 years; (2) confirmed ischemic stroke on MRI and CT within 21 days of symptom onset; (3) underwent OCTA imaging within 21 days of symptom onset; and (4) no presence of other neurological disorders, such as cerebral hemorrhage, Alzheimer's disease, or Parkinson's disease, etc. The exclusion criteria are as follows: (1) evidence of potential cardiac embolic sources (e.g. atrial fibrillation, recent myocardial infarction, valvular heart disease, dilated cardiomyopathy, and infective endocarditis); (2) contraindications to MRI; (3) inability to consent, or with cardiovascular diseases; (4) diagnosed with diabetic retinopathy or other retinal diseases; and (5) glaucoma. The assessment included lifestyle, National Institutes of Health Stroke Scale (NIHSS) score at admission, and clinical history (hypertension, diabetes mellitus, dyslipidemia, atrial fibrillation, coronary heart disease, smoking, and drinking) recorded at admission. For comparison, we enrolled individuals who attended our hospital but did not show any neurologic disorder from/through MRI as our control group. Individuals were excluded from the control group if they had the following: the presence of eye diseases, a history of ophthalmic surgery, toxic disorders affecting the central nervous system (CNS), a history of substance abuse, and medical ailment involving the use of concomitant corticosteroid or immunosuppressant therapy. Medical histories were recorded for all individuals. Lifestyle and clinical information were recorded for all controls. Questionnaires were used to assess age, sex, education level, and smoking status. Non–smokers and smokers were categorized based on their smoking status. Those who never smoked were classified as non–smokers, those who smoked regularly or had at least 100 cigarettes or 20 cigars in their lifetimes were classified as smokers, and those who actively consumed cigarettes were classified as smokers. Hypertension was defined as self–reported or use of antihypertensive medication or measured systolic blood pressure >140 mm Hg and/or diastolic blood pressure >90 mm Hg. A mercury sphygmomanometer was used to measure systolic and diastolic blood pressure, and their mean was calculated. Diabetes mellitus was defined as a self–reported diabetes history, use of anti–diabetic medication, or fasting blood glucose level >7.0 mmol/L. Dyslipidemia was defined as self–reported dyslipidemia history, use of dyslipidemia medication, or total cholesterol level of 6.21 mmol/L. 

To build a lacunar stroke and non–lacunar stroke classification cohort, patients with ischemic stroke presenting to the Neurology Department of West China Hospital of Sichuan University with first–ever stroke within 21 days of onset were recruited from December 2022 to August 2023. We classified the stroke subtype using radiologic criteria (whether the recent infarct on MRI was non–lacunar or lacunar) and used both the clinical and the radiologic classification to assign a final stroke subtype classification to the subject, as previously reported.¹⁵ 

OCTA Imaging

All used images in this study were acquired by swept–source OCTA (SS–OCTA; VG200S; SVision Imaging, Henan, China; version 2.1.016). As shown in Figure 1, the superficial vascular plexus (SVP), intermediate capillary plexus (ICP), and deep capillary plexus (DCP) were segmented by built–in software and these three en face angiograms were utilized in this study. SVP is described as the microvasculature between the base of the retinal nerve fiber layer (RNFL) to the junction between the inner plexiform layer (IPL) and inner nuclear layer (INL); ICP is defined as the vessels between the IPL/INL border to the border between the INL and the outer plexiform layer (OPL); DCP is defined as the microvasculature between the INL/OPL border to 25 µm below it. OCTA images covering an area of 3 × 3 mm² centered on the fovea were included in this study. Low–quality OCTA images (signal strength index [SSI] < 7/10) were excluded. 

Figure 1.

 

The definitions of OCTA stratification. In this study, SVP, ICP, and DCP en face images are included.

Deep Learning Model

We proposed an automated stroke identification model, namely ASI–Net, which is based on contrastive learning. The technical details of ASI–Net are shown in Figure 2. Specifically, we first used ConvNeXt¹⁶ as the backbone to extract the feature maps from the OCTA images. ConvNeXt uses a hierarchical feature extraction architecture and an attention mechanism that enables it to learn more complex representations of images. The feature maps with their corresponding labels are stored in the memory bank. Then, we followed the first–in–first–out (FIFO) strategy to update the memory bank following the stacking order, ensuring the reliability of these pseudo samples. Finally, we used these feature maps to calculate the similarity between these feature maps and other feature maps stored in the memory bank. This contrastive loss is used to force the model to focus on the differences between the common features of the two groups.^17,18 This is done by pulling together positive pairs and pushing apart negative pairs during training, based on the disease/control distinction. This allows the model to extract more fine–grained features, which improves the model's performance. 

Figure 2.

 

The workflow of our ASI–Net for stroke detection. The backbone was used to extract feature maps from the inputted OCTA en face images. The feature maps were also provided to the memory bank and used the contrastive learning method to learn the differences between the two different groups. Last, these features were sent to the classifier to output the prediction result.

To explore the performance of the given model, different retinal layers, including SVP, ICP, and DCP angiograms, were used as the input of the model, respectively. To train and evaluate the models, we used five–fold cross–validation, randomly dividing the dataset into a training set (80%) and a testing set (20%), with no overlap of patients. To further evaluate the generalization ability of our model, the model was tested on an independent dataset with a corresponding cohort. 

During training, random horizontal and vertical flip transformations were applied to each batch as data augmentation techniques to enhance and generalize the learning of the network. The Adam optimizer with a learning rate of 5 × 10⁻⁶ was utilized for training. The proposed model was implemented using Python and the PyTorch package, and all experiments were carried out on a workstation equipped with NVIDIA RTX 3090 graphics cards for training and validation. 

Statistical Analysis

The Shapiro–Wilk test was used to examine the normality of continuous variables. Characteristics were described as mean ± standard deviation (SD) or median [interquartile range] for continuous variables and frequencies with percentages for categorical variables. To examine the differences in the demographic and clinical variables between patients with ischemic stroke and the control group, the chi–square test was used for categorical variables. The t–test was used for normally distributed continuous variables, whereas the Mann–Whitney rank sum test was used for non–normally distributed continuous variables. 

Researchers have shown that aging and gender play a role in physiological changes. Moreover, vascular risk factors, such as hypertension, diabetes mellitus, dyslipidemia, and smoking, deteriorate the retinal vessels. To eliminate the interference of traditional risk factors, multivariable linear regression with a generalized estimating equation (GEE) was used to compare OCTA parameters between patients with ischemic stroke and control subjects. While adjusting for age and gender (model I) and subsequently adjusted by hypertension, hyperlipidemia, diabetes mellitus, smoking, and drinking (model II), P < 0.05 is considered statistically significant. Statistical analysis was performed with SPSS software package version 26.0 (SPSS. Inc., Chicago, IL, USA). 

Figure 3 shows the vascular parameters evaluated in this study. We investigate the vascular–related parameter representing the topology and shape of vessels, this being the vascular tortuosity (VT),¹⁹ and the foveal avascular zone circularity (FC).²⁰ All parameters were calculated using MATLAB version 9.4 (MathWorks. Inc, Natick, MA, USA). 

Figure 3.

 

Illustration results of vascular parameters. (a) This shows a 3 × 3 mm² en face SVP angiogram. (b) This is the detected FAZ area (FA), and (c) shows its perimeter (FP). The FAZ circularity is calculated as: FC = 4π · FA/FP ². (d) Illustrates the vascular branch map, it was used to segment vascular and calculate tortuosity.

Results

OCTA Dataset and Demographics

A STROBE flow diagram showing the exclusion criteria is presented in Figure 4. In the ischemic stroke identification cohort, we included 354 eyes from the 193 patients with ischemic stroke (median age = 58, IQR = 53 to 65, and 77.72% of men) and 331 eyes from 174 controls (median age = 58, IQR = 52 to 68, and 41.38% of men) as the internal dataset. As shown in Table 1, patients with ischemic stroke had a higher rate of smoking (P < 0.001), drinking (P < 0.001), hypertension (P < 0.001), dyslipidemia (P < 0.001), and diabetes (P = 0.003). An independent dataset from the Peking University Third Hospital was used to further evaluate the generalization ability of our model. This independent dataset includes 103 OCTA volumes of controls and 65 OCTA volumes of patients with ischemic stroke. OCTA images in this independent dataset were acquired using a spectral domain optical coherence tomography (SD–OCT) system (AngioVue, RTVue XR Avanti SD–OCT; Optovue, Fremont, CA, USA). In this independent dataset, each acquired OCTA volume corresponds to a single eye. 

Figure 4.

 

Flowchart for excluding participants who fail to meet the specified requirements. A total of 1730 eyes from 865 subjects were included in the study. IS denotes the ischemic stroke. I and T denotes the internal dataset and independent test dataset, respectively.

Table 1.

 

Demographic Characteristics of Two Internal Datasets

 

In the lacunar and non–lacunar stroke classification cohort, we included 260 eyes from the 142 patients with lacunar stroke (median age = 56, IQR = 51 to 63, and 71.83% of men) and 320 eyes from 171 patients with non–lacunar stroke (median age = 62, IQR = 54 to 69, and 78.95% of men) as the internal dataset. As shown in Table 1, there was no significant difference in gender, drinking, and dyslipidemia burden when both groups were compared. Patients with non–lacunar stroke are older (P < 0.001) and have a higher burden of hypertension (P = 0.012), furthermore, patients with non–lacunar stroke have a higher rate of smoking (P < 0.001) and diabetes (P = 0.018. An independent dataset from the Peking University Third Hospital was used to further evaluate the generalization ability of our model. It contains a total of 110 OCTA volumes, including 47 OCTA volumes of patients with lacunar stroke and 63 OCTA volumes of patients with non–lacunar stroke. OCTA images in this independent dataset were acquired using the –SS–OCTA tool (VG200S; SVision Imaging, Henan, China; version 2.1.016). In this independent dataset, each acquired OCTA volume corresponds to a single eye. 

Performance of Ischemic Stroke Identification

In this section, we displayed the performance of our deep learning method and compared it with other machine learning methods, including traditional machine learning methods and state–of–the–art deep learning classification models. 

Specifically, we compared ASI–Net with Logistic Regression Model, Gradient Boosting Tree, Support Vector Machine (SVM), VGG,²¹ ResNet,²² DenseNet,²³ EfficientNet,²⁴ and ConvNeXt. Table 2 shows the performance of the machine learning–based models in detecting patients with ischemic stroke using en face images. Of note, traditional machine learning methods, including Logistic Regression Model, Gradient Boosting Tree, and SVM, were trained using vascular parameters. ASI–Net achieved the highest performance in discriminating SVP images from the stroke and control groups in the 3 × 3 mm² area around the fovea, with an area under the receiver operating characteristic curve (AUC) of 0.922, an accuracy of 0.865, a sensitivity of 0.875, and a specificity of 0.855. 

Table 2.

 

OCTA Image–Based Stroke Identification Results Obtained by Different Machine Learning Models

 

The classification performances of ASI–Net using angiograms taken at different depths (SVP, ICP, and DCP) for training are shown in Table 3. Using 3 × 3 mm² SVP en face images, ASI–Net achieved the best performance (AUC = 0.922, accuracy = 0.865, sensitivity = 0.875, and specificity = 0.855). 

Table 3.

 

Classification Results of the Proposed Method When Using En Face Images for Training

 

We tested the model using the independent dataset to assess its generalization ability. As shown in Table 3, the model achieved an accepted performance (AUC = 0.822, accuracy = 0.732, sensitivity = 0.477, and specificity = 0.893). 

Performance of Lacunar and Non–Lacunar Stroke Classification

In this section, we showed the classification performance of ASI–Net using angiograms taken at different depths (SVP, ICP, and DCP) for distinguishing two stroke subtypes. As shown in Table 3, using 3 × 3 mm² SVP en face images, ASI–Net achieved the best performance (AUC = 0.871, accuracy = 0.784, sensitivity = 0.712, and specificity = 0.844). On the independent dataset, the best performance model achieved an accepted performance (AUC = 0.766, accuracy = 0.727, sensitivity = 0.630, and specificity = 0.721). 

Interpretability Analysis

We selected the best performance models of two cohorts to activate these heatmaps. Figure 5 demonstrates the heat maps of the given artificial intelligence (AI) models. In six randomly selected controls subjects and patients with stroke, it can be observed that the parafoveal region contributes the most to the classification decision –making of our AI model in both the stroke and control groups. In addition, the large range of blue colors in the heatmap also demonstrates that the AI can detect differences from a large area of vascularization. In six randomly selected patients with lacunar stroke and patients with non–lacunar stroke, the parafoveal region contributes the most to the classification decision –making of our AI model, in both the two stroke subtypes. Of note, the model activates more regions in the images of patients with lacunar stroke. 

Figure 5.

 

The heatmaps of the proposed model. (a) Presents the heatmaps of three randomly –selected healthy control (A–D) and three patients with stroke (D–F). (b) Presents the heatmaps of three randomly – selected patients with non–lacunar stroke (G–I) and three patients with lacunar stroke (J–L).

Statistical Analysis of Ischemic Stroke and Its Subtypes

Table 4 shows metrics of OCTA images of patients with ischemic stroke and controls in the internal dataset. After adjusting the traditional risk factors, patients with ischemic stroke showed increased SVP tortuosity (P = 0.037) and reduced SVP FAZ circularity (P = 0.043) when compared to controls; no significant differences (P > 0.05) in OCTA metrics were seen in the ICP and DCP when both groups were compared. 

Table 4.

 

Multivariable Linear Regression Analysis for Two Cohorts

 

The metrics of OCTA images of patients with lacunar stroke and patients with non–lacunar stroke in the internal dataset are also shown in Table 4. After adjusting the traditional risk factors, no significant differences (P > 0.05) were seen when the vascular tortuosity of the two groups was compared in angiograms taken at different depths. Patients with non–lacunar stroke showed reduced SVP FAZ circularity (P = 0.022) and reduced ICP FAZ circularity (P = 0.005) when compared to patients with lacunar stroke. No significant differences (P > 0.05) were seen in the DCP angiograms when the FAZ circularity of both groups was compared. 

Discussion

The proposed study introduces a retinal OCTA–based deep learning method, to detect microvascular changes in ischemic stroke and to classify ischemic stroke subtypes, that is, lacunar and non–lacunar strokes. Our carefully designed deep learning method showed promising performance in identifying retinal microvascular changes in ischemic stroke and classifying stroke subtypes from their OCTA images. Importantly, we showed that SVP angiogram had the highest distinct discrimination/discriminatory power to detect microvascular changes in ischemic stroke when compared with controls and also when classifying ischemic stroke subtypes. Individuals with ischemic stroke had increased retinal microvascular tortuosity and reduced FAZ circularity when compared with controls. 

Deep learning has been rapidly developing and is used to detect microvascular changes and also assist in the detection of many diseases. To the best of our knowledge, this study is the first to use a deep learning model in both distinguishing between the ischemic stroke and control groups and differentiating between the subtypes of ischemic stroke by using retinal OCTA images. Our ASI–Net model based on OCTA images demonstrated the highest performance of all the state–of–the–art methods tested, showing the efficiency and potential cost–effectiveness of the algorithm. 

Our method learns illness–related features from complex microvascular structures and eliminates the overfitting caused by the small size of the datasets through the introduction of contrastive learning. The results on the two independent datasets demonstrated our method's robust generalization ability. It is worth noting that, different from traditional machine learning algorithms that rely on manually extracted vascular parameters, deep learning methods are trained using en face images, simplifying the training process of the network. Importantly, with this deep learning architecture, our method has the potential to be applied to data from other medical centers, without developing a new deep learning model. Retrospective data can be collected from this specific medical center for supervised finetune, and the model can subsequently be refined to keep the deep learning model up to date. 

Vascular parameters, such as tortuosity and FAZ circularity, were introduced to study the correlation of ischemic stroke and its subtypes with retinal vascular changes. Traditional metrics, such as density and fractal dimension, can effectively evaluate the morphological changes of veins and venules. However, the complex microvascular structures in OCTA images hinder this scenario, reducing the sensitivity of these metrics to microvascular changes. In contrast, tortuosity can study the morphological changes of the vascular retina by examining the curvature of each vessel segment. Additionally, based on the interpretability analysis of deep learning models, we introduced FAZ circularity to the study. 

Of note, the SVP angiogram showed the highest discriminatory power to differentiate between ischemic stroke and controls. The SVP is a network of both large and small vessels connected directly to the retinal arteries and veins and supplies all other vascular plexuses. Several studies previously have suggested that cerebral small vessel disease is implicated in ischemic stroke.^25–28 Pathological mechanisms, such as atherosclerosis, led to the impairment of cerebral microcirculation highlighting the role of vessels in the disease cascade. Using the color fundus camera, reports have shown that individuals with ischemic stroke have wider retinal venules and narrower arterioles when compared with controls.^4,29 This implies that the retinal vasculature (arterioles and venules) is implicated during the stroke cascade, which reflects the cerebral microcirculation, as previously reported. Moreover, using quantitative measures on retinal imaging, recent reports have shown that patients with ischemic stroke have increased retinal vascular tortuosity when compared with controls.^30–32 Of note, the fundus camera displays the superficial vessel of the retina. Because the SVP contains the retinal arterioles, venules, and capillaries, increased SVP tortuosity in patients with ischemic stroke is in line with previous reports. Correspondingly, using our AI model, we showed that SVP angiograms had the highest discriminating power to detect microvascular changes between patients with ischemic stroke and healthy controls. Taken together, we suggest that SVP angiograms can be used to identify microvascular changes in patients with ischemic stroke and may constitute an early indicator to detect microvascular changes in ischemic stroke. Early identification of the microvascular changes in ischemic stroke may help clinicians to apply earlier implementation of treatment and may be potentially useful in predicting the progression of the microvascular complications related to ischemic stroke. 

We did not find any significant difference between the vessel tortuosity in the ICP and DCP when the ischemic stroke angiograms were compared to the controls. This discrepancy can be explained by understanding the effect of the increased hydrostatic pressure associated with ischemic stroke on different vessel sizes. It is suggested that a critical level of transmural pressure must be reached before tortuosity becomes evident.³³ Because the ICP and DCP are made of capillaries, it is plausible that the smaller microvasculature is more resilient and experiences this critical pressure in a more severe phase of the disease or longer duration. Taken together, we suggest that morphological changes in the SVP may be sensitive to structural changes that occur in ischemic stroke and may result from the downstream effect of cerebral microcirculation pathology. Thus, we suggest that the retina has the potential to be used as a screening tool to detect microvascular changes in patients with ischemic stroke. 

The causes of stroke are diverse, with multiple etiologies. Lacunar stroke is suggested to be due to cerebral small vessel disease, whereas non–lacunar stroke is suggested to be due to large vessels; thus, studying the retinal vascular changes in these two subtypes may provide valuable clues to their pathophysiology. Here, we showed that non–lacunar stroke had lower SVP and ICP FAZ circularity compared to lacunar stroke. The FAZ is suggested to be sensitive to ischemic changes and neurodegeneration. Although microvascular differences were not seen when both groups were compared, we suggest that neurodegeneration in lacunar stroke may be worse in non–lacunar stroke. 

We also explored the performance of using different en face images to train deep learning models, including SVP, ICP, DCP, inner–retinal vascular complex (IVC), and multi–inputs. In this study, deep learning models trained using various en face images demonstrated promising performance, including SVP, ICP, and DCP, suggesting that ischemic stroke and its subtypes may impact retinal vasculatures at different depths. Inspired by this, we explored the models’ performance in two cohorts by using the inner retinal vasculature. In each cohort, we used two methods to train the network: one using IVC en face images, and the other involving fusing SVP, ICP, and DCP en face images as inputs. As shown in Table 3, the models’ performance did not show significant improvement. This indicates that although IVC en face images contain more vascular information, it is limited by current imaging technologies, as mapping microvascular from different depths onto this 2D en face image may lead to overlapping and blurred vessel boundaries. This results in many unclear vascular details, which in turn limits vessel segmentation, parameter calculations, and the training of deep learning networks. Additionally, multi–input is widely used to enhance the performance of deep learning algorithms. In this study, we used the early fusion strategy by fusing SVP, ICP, and DCP en face images as inputs. However, the complex vasculature of the retina demands careful design of multi–input fusion methods. In the future, we will focus on designing specialized fusion modules to model the relationships between vascular structures in different en face images. 

This study has several limitations, the first of which stems from the relatively small dataset size compared with other studies in the deep learning field. Therefore, we used five–fold cross–validation to ensure the reliability of the model’s performance, and the carefully designed model achieved promising performance in the two specific tasks. Second, for the stroke identification model, the sensitivity is low (<50%) in the external validation dataset. This is mainly due to the problem of data heterogeneity. Our deep learning model was trained on the images acquired by SS–OCT (VG200S; SVision Imaging, Henan, China; version 2.1.016), where clear microvascular structures and subtle vascular changes in the images served as the basis for the classification of the network. Images in the validation dataset were acquired with an SD–OCT system (AngioVue, RTVue XR Avanti SD–OCT; Optovue, Fremont, CA, USA). In contrast, SD–OCT en face images show less clarity and detail of microvascular structures compared to SS–OCT. This limitation reduced the ability of the model to identify positive samples (ischemic stroke) based on microvascular structural changes. The differences in data distribution between the independent validation set and the training set may lead to reduced accuracy of the network in identifying positive samples. Last, to apply to clinical practice, the multi–disease dataset will be needed to validate the reliability of the model. As a preliminary study, further studies with multicenter, multi–ethnic, and multi–disease are planned to evaluate these AI algorithms. 

To our knowledge, this is the first study to apply deep learning on OCT angiograms to detect microvascular changes and to comprehensively assess quantitative measured retinal parameters in patients with ischemic stroke and its subtypes. Our study showed that SVP had the highest discriminating power to detect microvascular damage in ischemic stroke and its subtypes. Our study also provides detailed information about the degree and pattern of microvascular changes in ischemic stroke and its subtypes. We suggest that retinal microvasculature can reflect corresponding microvascular changes indicative of microvascular impairment leading to ischemic damage in the cerebral microcirculation. These novel results shown in our report have potential advantages beyond that of qualitative signs and quantitative parameters. 

Conclusions

In conclusion, our work demonstrated the applicability of artificial intelligence–enhanced OCTA image processing for identifying microvascular changes in ischemic stroke and its subtypes. Our AI analysis showed that the SVP–trained model had the highest performance in both ischemic stroke detection and its subtypes classification, and ischemic stroke individuals had increased vessel tortuosity in the SVP when compared with controls. These data provide convincing evidence that imaging of the retinal microvasculature using the OCTA has the potential to be used as a tool to detect microvascular changes in ischemic stroke. 

Acknowledgments

Supported in part by the National Science Foundation Program of China (62272444), in part by the Youth Innovation Promotion Association CAS (2021298), in part by the Zhejiang Provincial Natural Science Foundation of China (LR22F020008, LQ23F010007, LR24F010002, LZ23F010002), in part by Key research and development program of Zhejiang Province (2024C03101, 2024C03204) and Key Project of Ningbo Public Welfare Science and Technology (2023S012). 

Availability of Data and Materials: The data that support the findings of this study are available on request from the corresponding author. 

Disclosure: Z. Xiong, None; W.R. Kwapong, None; S. Liu, None; T. Chen, None; K. Xu, None; H. Mao, None; J. Hao, None; L. Cao, None; J. Liu, None; Y. Zheng, None; H. Wang, None; Y. Yan, None; C. Ye, None; B. Wu, None; H. Qi, None; Y. Zhao, None