Advancing Early Detection of Diabetic Retinopathy: A Novel Combo 4+1 Deep Learning Architecture with Classification by Regression

Diabetic retinopathy (DR) is a severe complication of diabetes, affecting over half of patients over time. This retinal disease often progresses without any symptoms in its early stages, then leading to irreversible blindness if left untreated. Accurate diagnosis and severity determination of DR are critical for successful treatment, but these steps are often time-consuming, costly, and requires expertise of trained ophthalmologists. The rise of artificial intelligence (AI) technology provides a paradigmshifting alternative to traditional diagnostic practices; still, previous studies have mainly focused on conventional classification schemes, often disregarding the ordinal nature that lies behind severity stages in diabetic retinopathy and the possibility of multi-scale feature extraction. The current study proposes a deep-learning (DL) architecture using ResNet50 transfer learning, demonstrating regression-based classification outperforming classical classification methods. This paper proposes a novel Combo <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$4+1$</tex> architecture, which divides each fundus image into its four quadrants and inputs them into the DL system, together with one downsized whole image. This design gives an effectively higher resolution for extracting features, allowing the detection of subtler lesions of neovascularization, while there is also an input of the whole image that gives a very different scale for extracting features. Using the APTOS 2019 dataset, the proposed model exhibits excellent results, particularly a binary classification accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 9. 0 5 \%}$</tex>, as well as a quadratic weighted kappa (QWK) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 3. 6 4 \%}$</tex> in five-class classification, which is a great improvement over existing results on the APTOS dataset. The proposed system has the potential to be a reliable screening mechanism and progression monitor of a patient's DR.