Innovative Fundus Image Grading and Categorization through Supervised Attention Multi-Scale Temporal Convolutional Network: A High-Precision Approach for Automated Fundus Image Quality Evaluation

The quality of retinal images generated by Computer-Aided Design (CAD) systems impacts their diagnostic performance for retinal disorders. Despite the extensive research on evaluating retinal image quality, a significant gap remains in accurately classifying gradable and non-gradable retinal images. In this Manuscript, fundus image gradeability categorization using a Supervised Attention Multi-Scale Temporal Convolutional Network (FIGC-SAMSTCN) is proposed. Input Images were gathered from the EyePACS dataset. Then, the images are fed into pre-processing utilizing the Robust Adaptive Smooth Variable Structure Kalman Filter (RASVSKF) to reduce the noise of the input image. The Pre-processed image is given to the Recursive Hilbert Transform (RHT) for extracting features like Contrast, Entropy, correlation, Energy and Homogeneity. Then the extracted features are provided to the Supervised Attention Multi-Scale Temporal Convolutional Network (SAMSTCN) to classify fundus image gradeability as gradable and ungradable. The Snow Ablation Optimizer (SAO) to achieve accurate classification. The proposed method attains 32.78%, 28.31%, and 14.56% higher Accuracy, 24.78%, 22.60% and 11.34% higher Precision, compared with existing techniques such as AFIP-DRG-FCN, IFIGC-CNN and the SFIF-CDR-LSTM. The lowest computational time of 2 ms is achieved by the proposed technique. The proposed technique is deployed in automated retinal screening systems and clinical decision support platforms to enable rapid assessment of fundus image quality facilitate early detection of eye diseases and improve the efficiency and accessibility of ophthalmic care.