Cardiovascular Disease Risk Assessment from Retinal Fundus Images Using a CSP-ConvNext Model

Abstract

Background/Objectives: Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, making early identification of individuals at elevated cardiovascular risk essential for preventing major adverse cardiac events (MACE). Retinal fundus photographs provide a non-invasive window into systemic vascular health, as retinal microvascular and optic-disc characteristics have been associated with cardiovascular abnormalities. This study aims to develop an enhanced ConvNeXt-based deep learning framework, termed CSP-ConvNeXt, for automated prediction of CVD-related risk categories from retinal fundus images, including Mild, Moderate, Severe, and Proliferative Diabetic Retinopathy (DR), as well as Hypertensive Retinopathy (HR). Methods: The proposed CSP-ConvNeXt framework integrates a hierarchical convolutional backbone with Cross Stage Partial (CSP) channel partitioning and large-kernel depthwise convolutions to capture both fine-grained retinal microvascular features and broader anatomical structures associated with cardiovascular risk. An Efficient Squeeze-and-Excitation (ESE) attention mechanism is incorporated to enhance the representation of clinically relevant vascular regions while reducing the influence of irrelevant background information. A robust data-augmentation and preprocessing pipeline is employed to improve model generalizability under varying retinal imaging conditions. Two benchmark retinal datasets, comprising Mendeley Retinal Fundus images and the APTOS 2019 dataset, are used for model training and evaluation. Results: The CSP-ConvNeXt framework achieves an overall classification accuracy of 87%, demonstrating effective discrimination among the considered CVD-related retinal risk categories. The model achieves high precision and recall values, indicating consistent classification performance across the evaluated categories. Statistical analysis demonstrates significant interclass differentiation, with a p-value < 1 × 10−16, supporting the discriminative capability of the learned retinal representations. Conclusions: The results demonstrate that the CSP-ConvNeXt framework, combining CSP-based feature extraction, ESE attention, large-kernel depthwise convolutions, and hierarchical representation learning, can effectively capture retinal vascular and anatomical patterns associated with cardiovascular risk. The results highlight the potential of advanced ConvNeXt-based architectures for automated, non-invasive CVD-related risk screening using retinal fundus photographs. This approach may provide a valuable complementary tool for early cardiovascular risk assessment, particularly in settings where conventional cardiovascular screening is less accessible.