Multi-Label Classification of Fundus Images with Optimized Vision Transformer Based Wasserstein Deep Convolutional Generative Adversarial Network

• Published in: SN computer science 2024-08, Vol.5 (7), p.842, Article 842
• Main Authors: Sreekumar, K., Vimina, E. R.
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Blood vessels; Classification; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Datasets; Deep learning; Design of experiments; Diabetes; Diagnosis; Disorders; Eye diseases; Feature extraction; Fourier transforms; Generative adversarial networks; Glaucoma; Illnesses; Image enhancement; Image filters; Image quality; Image restoration; Information Systems and Communication Service; Labels; Literature reviews; Machine learning; Medical imaging; Medical personnel; Multivariate analysis; Neighborhoods; Neural networks; Optic nerve; Original Research; Parameters; Pattern Recognition and Graphics; Pyramids; Retina; Software Engineering/Programming and Operating Systems; Tortuosity; Vision
• ISSN: 2661-8907; 2662-995X; 2661-8907
• DOI: 10.1007/s42979-024-03161-0

Early retinal disease diagnosis and treatment are essential for preventing the irreversible vision impairment. The patients in the clinical settings have various kinds of retinal disorders. Fundus image categorization is a multiple label categorization task since a fundus image contains one or more disorders. However, most previous research has focused on the diagnosis of a single fundus problem and there remain substantial obstacles to the simultaneous diagnosis of several fundus conditions. In this research work, Multi-Label Classification of Fundus Images with Optimized vision transformer based Wasserstein Deep Convolutional Generative Adversarial Network (MLC-FI-VF-WDCGAN) is proposed to accurately classify the Fundus disease images. Initially, the input image data is collected from Ocular Disease Intelligent Recognition (ODIR) dataset. The input images are preprocessed with Multivariate Fast Iterative Filtering (MFIF) for eliminating the noise and increasing the quality of input imageries. The pre-processing images are given to Multi-scale Neighborhood Feature Extraction (MS-NFE) method for extracting features such as, Geometric Features, Blood Vessel and Vascular Tortuosity for enhancing the classification accuracy. Then, the extracted features are provided to vision transformer based Wasserstein Deep Convolutional Generative Adversarial Network (VF-WDCGAN) for classifying the Fundus disease as Normal, Diabetic retinopathy, Glaucoma, Cataract, Hypertensive retinopathy and Myopia. The Vision-transformer acts as Discriminator in the WDCGAN and doesn’t reveal any adoption of optimization methods for calculating the ideal parameters in improved categorization accuracy. Hence, Giza Pyramids Construction Algorithm (GPCA) is proposed to improve the weight parameters of VF-WDCGAN. The MLC-FI-VF-WDCGAN technique is implemented in python and evaluated using several performances metrics such as accuracy, precision, recall. The performance of proposed MLC-FI-VF-WDCGAN method is compared with existing techniques like, deep variational auto-encoders for denoising retinal fundus image (DVAE-DR-FI), Cycle GAN-depend deep learning method for artifact reduction in fundus photography (CGAN-DLAR-FP) and Rformer: Transformer-depend generative adversarial network for real fundus image restoration on new clinical benchmark (TGAN-FI-NCB) respectively.