The proposed DD-FIF-CDI-VCNN framework, combining fundus imaging features with clinical demographic information via a Verifiable Convolutional Neural Network optimized by the Warthog Optimization Algorithm, achieves enhanced diabetic retinopathy detection performance compared to existing unimodal approaches.
The DD-FIF-CDI-VCNN method achieved superior performance across multiple evaluation metrics compared to existing models for diabetic retinopathy detection.
Robust Consensus Tobit Kalman Filtering (RCTKF) was used as a preprocessing method to address image quality issues including low contrast, noise, and pixel intensity normalization.
A ResNet-fused External Attention Network (ResfEANet) was used to extract rich visual and texture features from retinal fundus images.
TabNet was used to extract meaningful feature vectors from structured demographic and clinical attributes of patients.
Hierarchical Multi-Scale Feature Fusion (HMSFF) was employed to integrate visual, texture, and demographic features into a unified representation.
The Verifiable Convolutional Neural Network (VCNN) classifier parameters were optimized using the Warthog Optimization Algorithm (WOA) to improve convergence and detection performance.
The proposed framework incorporated bilateral eye correlation by analyzing fundus images from both eyes, addressing a limitation of traditional single-eye analysis systems.
The proposed framework addressed multiple known limitations of existing diabetic retinopathy detection systems including class imbalance, low image contrast, and single-disease focus.
This research suggests that combining eye scan images with patient health information (such as demographics and clinical data) in a single AI system can improve the detection of diabetic retinopathy, a leading cause of blindness in people with diabetes. The proposed system, called DD-FIF-CDI-VCNN, processes fundus photographs (images of the back of the eye) from both eyes simultaneously — rather than one eye at a time as most existing systems do — and merges this visual information with patient-specific data using a series of specialized AI components. The system also applies image enhancement techniques before analysis to reduce noise and improve image quality, addressing common problems in real-world medical image datasets. The framework uses several layered AI techniques: a neural network called ResfEANet to extract visual patterns from the eye images, a tool called TabNet to process structured patient data, a fusion method called HMSFF to combine all information sources, and a final classifier called VCNN whose settings are automatically tuned using an optimization algorithm inspired by warthog behavior. Together, these components work to classify patients as having normal eyes or diabetic retinopathy. The researchers compared this approach against existing models using standard performance measures including accuracy, precision, recall, F1-score, AUC, and error rate. This research suggests that incorporating patient demographic and clinical information alongside eye imaging — rather than relying on images alone — can make AI-based diabetic retinopathy screening more accurate and robust. If validated more broadly, such multimodal approaches could help clinicians catch diabetic eye disease earlier and more reliably, potentially reducing the risk of vision loss in people with diabetes. However, as this study relies on a single publicly available dataset, further validation across diverse real-world clinical settings would be important before widespread adoption.
Mohammad A, Ramakrishnaiah N. (2026). Enhanced diabetic retinopathy detection via multimodal fusion of fundus imaging features and clinical demographic information using verifiable convolutional neural network.. Computers in biology and medicine. https://doi.org/10.1016/j.compbiomed.2026.111892