Multi-Scale ODE-Based Continuous Feature Extraction for Retinal Disease Classification Using Fundus Images

Abstract

The automated diagnosis of retinal diseases from fundus images has attracted a lot of attention due to the increasing prevalence of vision-threatening ocular disorders and shortcomings associated with manual clinical screening. However, multiclass classification of retinal diseases is inherently difficult due to inter-class similarities, changing illumination conditions, low-contrast structures in the retina and computationally expedient models are preferred. In this paper, a new deep learning framework is proposed to classify multiclass retinal diseases based on fundus images. In order to enhance the ability in effectively resulting a desired output for retinal feature representation, this research presents a method that integrates multi-scale feature extraction with an ODE-based learning strategy. The framework automatically learns hierarchical retinal representations from raw fundus images using CNN and multi-scale ODE modules, thereby avoiding hand-crafted feature engineering and enabling end-to-end disease classification. Experiments are performed on the publicly available Eye Disease Image Dataset which consists of images related to multiple categories of retinal diseases. Various preprocessing, data augmentation and feature scaling techniques were incorporated in the pipeline to ensure robustness and generalization. The proposed framework was developed and evaluated for the classification of 10 retinal disease categories using retinal fundus images from the Eye Disease Image Dataset (EDID). The proposed ODE based approach successfully classifies multi-class retinal disease, and the experimental results confirm its performance. Under this framework, ODE-driven feature evolution is used to gradually update the representation of the retinal image by modeling how edges, textures, shapes and pathological patterns in fundus images change over time. The method achieved a classification accuracy of 86.40% and macro precision, sensitivity, F1-score, specificity and balanced accuracy were measured as 87.92%, 88.62%, 88.24%, 98.41% and 88.61%, respectively. Furthermore, three levels of visualization analysis are provided, including ODE feature vector visualization, ODE overlay and Grad-CAM interpretability analysis, and suspicious region localization. The visualizations corroborate that the proposed method achieves a balance of separation of discriminative pathological features while maintaining interpretability, accuracy and accurate localization making it applicable for robust retinal disease diagnosis.