Compact Pyramidal dense mixed attention network for Diabetic retinopathy severity prediction under deep learning

• Published in: Biomedical signal processing and control 2025-02, Vol.100, p.106960, Article 106960
• Main Authors: Gargi, M., Krishna Eluri, Rama, Prakash Samantray, Om, Hajarathaiah, Koduru
• Format: Article
• Language: English
• Subjects: Axial spectral transformer; Equilibrium optimization algorithm; Fundus images; Non-local means Gaussian filter; Pyramidal dense mixed attention network
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106960

•Noise removal using the CE-NLG to minimize the error and processing time.•To extract the features, E-ASpecT is used that minimizing the model time complexity.•The selection of features is done using the DuSEO for reducing the convergence issues.•To attain DR severity classification, a novel DL model called CPDenA is introduced. Diabetic Retinopathy is the leading cause of blindness in people with diabetes. Artificial intelligence and camera technology advancements have enabled the ability to autonomously diagnose the severity of DR. Nevertheless, the existing approaches target only to enhance the classification task as presence and absence. However, the severity classification remains a difficult issue that must be achieved in order to overcome the limitations of present studies. In this paper, a unique DR severity prediction method based on Deep Learning algorithms is proposed. For ophthalmologists, digital fundus images are the most routinely used imaging for screening and determining disease severity. The fundus images are first collected from the Diabetic Retinopathy Dataset and Asia Pacific Tele-Ophthalmology Society (APTOS-2019) datasets and then pre-processed using the Contrast-enhanced Non-local means Gaussian filter to remove noise, improve contrast, and sharpen the image. To reduce time complexity and improve the model’s ability, the features are extracted using the Excess level attention axial spectral transformer. To reduce convergence and dimensionality concerns, an ideal collection of features is chosen from the retrieved features using the Dual-phase sine equilibrium optimization algorithm. The severity of the DR can be efficiently characterized across multiple classes utilizing the Compact Pyramidal dense mixed attention network model. The CPDenA model is evaluated using numerous DL performance measures such asaccuracy, recall, precision, etc. In the APTOS-2019 and DDR datasets, the CPDenA model achieved high accuracy of 97.95% and 98.59%, respectively. When compared to other previous research, the CPDenA model outperformed in predicting DR severity.