Identification of cardiovascular diseases from colour fundus images using ensemble texture features by deep neural networks

Cardiovascular diseases or CVDs (such as arrhythmia, myocardial infarction, heart failure, stroke, hypertension, and coronary artery diseases) are among the leading causes of early mortality worldwide, particularly in low- and middle-income populations due to changing lifestyles and associated health conditions. Therefore, early prediction of CVD risk is essential for timely diagnosis and effective clinical management. CVD risk can be assessed using modalities such as electrocardiogram (ECG), phonocardiogram (PCG), and retinal fundus imaging. Among these, fundus images provide non-invasive visualization of retinal vascular structures and associated cardiovascular abnormalities. However, accurate prediction from fundus images remains challenging due to variations in retinal texture, illumination, vascular morphology, and disease-specific structural patterns. This manuscript uses local binary pattern (LBP) and gray-level co-occurrence matrix (GLCM) features to capture detailed texture information from colour fundus images and ensembles them with deep neural network (DNN) models for CVD classification. LBP captures local texture patterns, while GLCM represents global retinal texture characteristics. The ensemble of LBP and GLCM features enables extraction of both fine and coarse retinal textures associated with CVD risk. Three deep learning (DL) architectures, namely DenseNet-121, EfficientNet-B3, and MobileNetV3-Large, were used for classification. The proposed technique was evaluated on the China-Fundus-CIMT dataset, which consists of two groups based on carotid intima media thickness (CIMT). The proposed framework was evaluated using standard evaluation metrics including accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), and AUC. DenseNet-121 achieved the highest accuracy of 73.30% with an MCC of 0.389, when applied over the images directly. Ensemble texture features (i.e., LBP + GLCM) obtained the best overall performance (i.e., accuracy of 75.20%, F1-score of 71.00%, and MCC of 0.422) on DenseNet-121. The ensemble of LBP and GLCM features improved the classification accuracy of MobileNetV3-Large from 69.40% to 72.90%, demonstrating the effectiveness of applying multiple handcrafted texture features on DL based classification. The suggested framework improves CVD classification performance by integrating handcrafted retinal texture descriptors with computationally efficient DL models. Unlike several existing fundus image-based CVD prediction approaches, it does not rely on additional clinical risk factors, multimodal data, or computationally intensive architectures.