Deep Learning Based Estimation of Blood Glucose Levels from Multidirectional Scleral Blood Vessel Imaging
Authors: Muhammad Ahmed Khan, Manqiang Peng, Ding Lin, Saif Ur Rehman Khan
Organizations: Changsha Aier Eye Hospital, South Furong Rd, Changsha, Hunan, China. · Aier Eye Hospital Group Co., Ltd, South Furong Rd, Changsha, Hunan, China. · Central South University AIER Eye Institute of Ophthalmology, Changsha, Hunan, China. · Central South University, Changsha, Hunan, China. · German Research Center for Artificial Intelligence, Kaiserslautern, 67663, Germany.
Regular monitoring of glycemic status is essential for diabetes management, yet conventional blood-based testing can be burdensome for frequent assessment. The sclera contains superficial microvasculature that may exhibit diabetes related alterations and is readily visible on the ocular surface. We propose ScleraGluNet, a multiview deep-learning framework for three-class metabolic status classification (normal, controlled diabetes, and high-glucose diabetes) and continuous fasting plasma glucose (FPG) estimation from multidirectional scleral vessel images. The dataset comprised 445 participants (150/140/155) and 2,225 anterior-segment images acquired from five gaze directions per participant. After vascular enhancement, features were extracted using parallel convolutional branches, refined with Manta Ray Foraging Optimization (MRFO), and fused via transformer-based cross-view attention. Performance was evaluated using subject-wise five-fold cross-validation, with all images from each participant assigned to the same fold. ScleraGluNet achieved 93.8% overall accuracy, with one-vs-rest AUCs of 0.971,0.956, and 0.982 for normal, controlled diabetes, and high-glucose diabetes, respectively. For FPG estimation, the model achieved MAE = 6.42 mg/dL and RMSE = 7.91 mg/dL, with strong correlation to laboratory measurements (r = 0.983; R2 = 0.966). Bland Altman analysis showed a mean bias of +1.45 mg/dL with 95% limits of agreement from -8.33 to +11.23$ mg/dL. These results support multidirectional scleral vessel imaging with multiview learning as a promising noninvasive approach for glycemic assessment, warranting multicenter validation before clinical deployment.
Figures & tables
Figure 1: Overview of the study workflow
Group
Subjects (n)
Images per subject
Total images
Normal
150
5
750
Controlled diabetic
140
5
700
Uncontrolled diabetic
155
5
775
Total
445
–
2225
Table 1: Dataset formation. Range of participants and related multidirectional scleral photos from the normal, regulated diabetic, and unregulated diabetic groups. Each participant submitted five images of the anterior segment, resulting in a total of 2,225 images.
Parameter
Normal
Controlled diabetic
High Glucose diabetic
Age (years)
41.8±11.4
48.3±10.8
53.6±9.7
HbA1c (%)
5.3±0.3
6.9±0.4
9.1±1.2
Fasting plasma glucose (mg/dL)
92.4±8.1
133.8±14.7
204.9±28.6
Gender (M/F)
82 / 68
76 / 64
88 / 67
Table 2: Clinical and demographic characteristics of the study population. Values are reported as mean ± standard deviation. FPG: fasting plasma glucose; HbA1c: glycated hemoglobin.
Figure 2: Outlines the protocol for the standardized form of scleral imaging. To check focus and alignment of the images, the first image taken is of a straight gaze position. This is followed by the capture of four other gaze direction images which are necessary for assessing all of the scleral regions. These images are taken in the upward, downward, leftward, and rightward gaze positions, respectively, so that the inferior, superior, temporal, and nasal scleral regions are photographed. By using this protocol, all slices of the scleral vasculature can be photographed for analysis from multiple views.
Figure 3: Procedure of image pretreatment and enhancement of blood vessels. This image shows the anterior segment with raw data, scleral portion that has been cropped and normalized, an image that has vessels that have been enhanced using CLAHE, and a multi scale Frangi filter, and the related binary image of the blood vessels that has been segmented and that has been used to visualize and control the quality of the image.
Figure 5: The Proposed ScleraGluNet and the Previous MRFO-INEYENET model. Despite the fact that MRFO-INEYENET relies on a single glance of ocular pictures and on the MRFO-based feature optimization, ScleraGluNet advances this system by integrating multi directional scleral imaging, and by also adding parallel, branch wise, view specific convolution and transformer based cross view fusion to aid in the simultaneous realization of classification and regression.
Category
Participants
Scleral Images
Notes
Control
150
750
Five images per participant
Controlled Diabetic
140
700
Five images per participant
Uncontrolled Diabetic
155
775
Five images per participant
Total
445
2225
–
Table 3: Participants’ distribution and scleral imaging across metabolic categories
Characteristic
Control
Controlled Diabetic
Uncontrolled Diabetic
Age (mean ± SD)
–
–
–
Sex (M/F)
–
–
–
FPG (mg/dL)
–
–
–
HbA1c (%)
–
–
–
Table 4: Participants’ demographics and clinical information
Class
Precision (95% CI)
Recall (95% CI)
F1-score (95% CI)
Normal
0.934 (0.882–0.964)
0.940 (0.890–0.968)
0.937 (0.898–0.957)
Controlled diabetic
0.915 (0.857–0.951)
0.921 (0.865–0.956)
0.918 (0.874–0.943)
High Glucose diabetic
0.948 (0.900–0.973)
0.935 (0.885–0.965)
0.942 (0.904–0.961)
Macro-average
0.93
0.93
0.93
Weighted average
0.93
0.93
0.93
Table 5: Class-wise performance of ScleraGluNet using five-fold cross-validation. Precision, recall, and F1-score (with 95% CIs) are presented for each class, along with macro and weighted averages.
Figure 6: Confusion Matrix for Three-Class Metabolic Status Classification. The rows show the true class labels, and the columns show the predicted class labels. The values present the subject counts integrated over the five-fold cross-validation. Counts represent participants (not images).
Figure 7: The one-vs-rest ROC curves are for the three-class metabolic status ROC curves. They are for the normal (AUC = 0.971), controlled diabetes (AUC = 0.956), and High Glucose diabetic (AUC = 0.982). These ROC curves demonstrate the discrimination of the proposed model across different thresholds.
Figure 8: Regression results from a plasma glucose estimation. A chart illustrating congruence of predicted FPG values with lab verified results. The predictive model accuracy was statistically stellar (r = 0.983). Furthermore, they registered a root mean square of 0.44 [7.91] indicating an almost faultless glucose estimation through the different ranges of glycemic levels.
Figure 9: Bland–Altman analysis for FPG estimation. The graph shows the difference between FPG measured in the laboratory and predicted FPG values their mean FPG values. The mean bias was 0.08 mmol/L (1.45 mg/dL) and the 95% limits of agreement -0.46 to +0.62 mmol/L. This shows that the proposed method and the reference measurements agree well throughout the range of FPG values.
Fold
Accuracy (%)
Macro-Precision
Macro-Recall
Macro-F1
1
93.2
0.92
0.93
0.93
2
94.1
0.94
0.94
0.94
3
92.8
0.92
0.92
0.92
4
94.6
0.94
0.94
0.94
5
93.9
0.93
0.94
0.94
Mean ± SD
93.7 ± 0.7
0.93 ± 0.01
0.93 ± 0.01
0.93 ± 0.01
Table 6: Fold-wise classification performance of ScleraGluNet in the five-fold cross validation experiment. High accuracy across all folds indicates the stability of the framework, with little variance showing that performance is not dependent on a favorable data split.
Figure 10: The Scleral images and their related explainability maps in different metabolic phases are shown. Each row represents a subject in one of three states: normoglycemic, controlled diabetic, and poorly controlled hyperglycemic. Each column includes one original scleral image along with the corresponding Grad-CAM and Grad-CAM++ visualizations. The heatmaps in the visualizations show the regions of the images that have the most significant impact on the classification decision, and these areas of highest influence are indicated in red.
Figure 11: The Scleral images that were collected while looking in different directions and during different metabolic conditions are shown using Grad-CAM visualizations. The subjects in each row have normal glucose levels, controlled diabetes, and high glucose levels that are not managed, in that specific order. The images in each row are organized based on gaze positions which are vertical, downward, leftward, and rightward. The Grad-CAM overlays function as heatmaps, indicating areas in the images that contribute more significantly by using warm colors placed over the original scleral images.
Figure 12: Images of sclera and glucose projections according to metabolic categories. For every individual, particular multiple-angle scleral photographs and laboratory-recorded along with model-calculated values of fasting glucose levels are exhibited. In normal instances, there are regular and fine blood vessels, while there are mild-to-moderate irregularities of blood vessels in controlled diabetes, and in High Glucose diabetic, there are high degrees of vascular abnormalities. In these cases, the estimated glucose levels highly correlated with the measured values.
Figure 13: The Multiview ScleraGluNet CNN architecture MAE in mmol/L is derived from an ablation study on the models, which include the single-view CNN baseline, the multiview CNN without MRFO, the multiview CNN with MRFO, and the complete ScleraGluNet. The findings show that the performance enhancement brought by multiview learning, optimization of features, and MRFO-based transformer fusion is incremental.
Diabetic retinopathy screening traditionally relies on fundus photography, requiring specialized equipment and expertise often unavailable in primary care and resource limited settings. We developed and validated a deep learning (DL) system for automated diabetic classification using anterior segment ocular imaging a readily accessible alternative utilizing standard photography equipment. The system leverages visible biomarkers in the iris, sclera, and conjunctiva that correlate with systemic diabetic status. We systematically evaluated five contemporary architectures (EfficientNet-V2-S with self-supervised learning (SSL), Vision Transformer, Swin Transformer, ConvNeXt-Base, and ResNet-50) on 2,640 clinically annotated anterior segment images spanning Normal, Controlled Diabetic, and Uncontrolled Diabetic categories. A tailored preprocessing pipeline combining specular reflection mitigation and contrast limited adaptive histogram equalization (CLAHE) was implemented to enhance subtle vascular and textural patterns critical for classification. SSL using SimCLR on domain specific ocular images substantially improved model performance.EfficientNet-V2-S with SSL achieved optimal performance with an F1-score of 98.21%, precision of 97.90%, and recall of 98.55% a substantial improvement over ImageNet only initialization (94.63% F1). Notably, the model attained near perfect precision (100%) for Normal classification, critical for minimizing unnecessary clinical referrals.
Hasaan Maqsood, Saif Ur Rehman Khan, Sebastian Vollmer +2
BiogentX (biogentx.com, Renala Khurd, District Okara, Punjab, Pakistan. · Department of Computer Science, Rhineland-Palatinate Technical University of Kaiserslautern-Landau, Kaiserslautern, 67663, Germany. · German Research Center for Artificial Intelligence, Kaiserslautern, 67663, Germany. +2
Glaucoma is a progressive eye disease that can lead to irreversible vision loss if not detected at an early stage. Conventional diagnostic procedures are often time-consuming and rely heavily on expert interpretation, limiting their scalability for large-scale screening. In this study, glaucoma detection is investigated under two evaluation settings: sample-wise, where individual samples are analyzed independently, and patient-wise, where data from each patient are aggregated for final prediction. An automated multimodal framework is proposed that integrates fundus images with clinical data. Under the sample-wise setting, detection is performed using fundus images and clinical features individually, as well as through their multimodal combination. Under the patient-wise setting, predictions are obtained by aggregating multiple fundus image representations with corresponding clinical information for each patient. Deep visual features are extracted using a Vision Transformer (ViT) architecture and classified using classical machine-learning models, with a stacking-based ensemble of the three best-performing classifiers employed to optimize performance. Experiments conducted on the publicly available PAPILA dataset demonstrate strong diagnostic performance, achieving 97.47% accuracy and a 97.50% F1-score for sample-wise multimodal classification, and 98.97% accuracy and F1-score for subject-wise detection. The proposed framework is further deployed as an end-to-end web-based platform to support automated glaucoma screening and clinical decision support.
Ishrat Jahan, Muhammad E. H Chowdhury, Murugappan Murugappan +7
Department of Computer Science and Engineering, Shahjalal University of Science and Technology, Sylhet, Bangladesh · Department of Electrical Engineering, Qatar University, Doha, Qatar · Department of Electronics and Communication Engineering, Kuwait College of Science and Technology, Block 4, Doha, Kuwait +5
We propose glaucoma lesion evaluation and analysis with multimodal imaging (GLEAM), the first publicly available tri-modal glaucoma dataset comprising scanning laser ophthalmoscopy fundus images, circumpapillary OCT images, and visual field pattern deviation maps, annotated with four disease stages, enabling effective exploitation of multimodal complementary information and facilitating accurate diagnosis and treatment across disease stages. To effectively integrate cross-modal information, we propose hierarchical attentive masked modeling (HAMM) for multimodal glaucoma classification. Our framework employs hierarchical attentive encoders and light decoders to focus cross-modal representation learning on the encoder.
Jiao Wang, Chi Liu, Yiying Zhang +8
College of the Information Science and Engineering, Northeastern University, Shenyang 110819, China · Department of Ophthalmology, Shenyang Fourth People’s Hospital, Shenyang 110000, China