GAN-DCNN: A Generative Adversial Network based Deep Convolutional Neural Network for Retinal Vessel Enhancement

GAN-DCNN: A Generative Adversial Network based Deep Convolutional Neural Network for Retinal Vessel Enhancement | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article GAN-DCNN: A Generative Adversial Network based Deep Convolutional Neural Network for Retinal Vessel Enhancement S. Arun Kumar, M. Karuppasamy, T. Subburaj, David Neels Ponkumar, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7017188/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Artificial intelligence is taking over the technology transmission in all areas. Generative AI is now mostly used in the healthcare sector due to its vivid application area. The emergence of Generative Adversial Network (GAN) in medical domain is highly recommended. The proposed work brings the essence of deep convolutional network that differentiates real and generated retinal images. This work highlights the advancement of GAN over other artificial intelligence models. By identifying appropriate networks of GAN, retinal images are generated in high quality. Low resolution images are enhanced with GAN models. The proposed work involves basic preprocessing, augmentation and enhancement with the retinal images that provide better accuracy with epochs. Deep learning models with GAN implementation are highly preferred for the retinal classification. Generative AI generative adversial network healthcare retinal images disease prediction DCGAN SRGAN Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1 Introduction Artificial intelligence has become the most popular tool in various application areas which include healthcare sector that helps in clinical decision making and improving health outcomes [ 1 ]. Generative AI is one of the AI area which generates new data and images that assist patient care, providing clinical decision making [ 2 ]. In the present era, Generative Adversial Networks have evolved as a vital tool in the medical imaging and healthcare data analytics. It is used in variety of applications like medical image synthesis, data augmentation, anamoly detection and disease diagnosis [ 3 , 4 , & 5 ]. The indepth training mechanism empoers GAN to observe intricate data distribution and produce quality that helps in training data. With the use of various medical images, GAN generates synthetic images with diverse variance and also helps in facilitating the augmentation of training data. With advanced deep learning models GAN also helps in better prediction if diseses especially in medical imaging. GAN generates realistic synthetic images that include pathological images which contributes to the study of rare diseases and abnormalities [ 6 ]. GAN plays a diverse role in medical image enhancement and reconstruction [ 7 ]. GAN methodologies convert nosiy images into high quality images that enhances the procedures and treatment strategies. GAN are used in various health related issues like cardiovascular diseases, diabetes, cancer and medical image analysis [ 9 ]. Figure 1 shows the application areas of GAN in medical and healthcare sectors. GAN uses the training of two neural networks for image generation and discrimination. The generative model uses training data to produce images while discriminative model estimates the image came from training data. Augmentation Augmentation is majorly used in data and image processing that helps in medical image analysis where there is limited dataset [ 10 ]. This involves various features like rotation, scaling and flipping to elaborate the dataset size. Incase of medical imaging there will be shortage of images to train a mode where there will be less accurate results, while augmentation helps in overcoming this issue. Augmentation expands the dataset size required for efficient training and provide more robust and accurate results [ 11 ]. Segmentation Segmentation in medical images is used to improve the image quality. This also allows multi-modal fusion, helps in image reconstruction and enhance the medical image analysis. Medical diagnoses are more perfect with reliable results when segmentation techniques are involved [ 12 ]. Reconstruction It is a process that helps in reconstructing an image with the help of algorithm that builds image from an incomplete data. This is mostly used in medical imaging, remote sensing and computer vision. This process usually generated high-fidelity images from low resolution images [ 13 ]. This also removes noise using generative adversial networks. GAN consists of generator and discriminator. Generator learns to real images from noisy or imperfect images. In particular with medical image analysis, GAN generated images are valuable like super-resolution images with enhanced image quality. This assist the medical sector in analyzing medical scans with clear and detailed images for accurate results. Detection GAN has become a vital tool in medical image analysis. This also helps in detection of diseases or medical conditions such as cancer, retinal abnormalities, Alzhemier’s and other anomalies. GAN generates synthetic images of real data and compares with the real medical images. This enhances the accuracy of early disease detection [ 14 ]. If diagnosis is done earlier there is high possibility of reduction of risk in major diseases. Classification GAN is an indirect mechanism for classification by diversifying training data through data augmentation. Though GAN does not directly support classification, it is a hidden factor that enhances feature extraction, detect anomalies and also provide test samples in case of semi-supervised learning. With these process the classification accuracy is much improved and utilized for various applications. Deep learning based medical image classification is also improved with GAN [ 15 ]. Image Synthesis GAN is used to generate synthetic medical images that represents similar characteristics of patient data. It contains generator that generates images and discriminator that differentiate between real and synthetic images. In healthcare sectors, medical image synthesis enables training of the machine learning models. With the help of data or image generation the size of the dataset is also increased making it easier for training machine learning models, at the same time the privacy of the data is also conserved. GAI is specially more helpful for personalized treatment that involves predictive modeling to featuring treatment plans. Conditional Variant Encoder combines conditional variables and autoencoders to learn patient information that incorporated specific conditions for treatment plans. By specifying different conditions the model could generate various treatment strategies to individual patients. Fine Tuned NLP such as BERT are improved in generating personlaized plans for patient information. GAI with GAN are almost used for medical image analysis. It is also helpful in monitoring various health conditions, diabetes analysis, medical image analysis and cancer detection. 2 Related Work There were several applications of GAN in retinal image classification that include segmentation, data augmentation, producing high resolution images, prediction and feature extraction. Perception oriented GAN for retinal images that provide high resolution images are employed [ 16 ]. Another study employed adversial technique on vascular network with some concerns overcome by training autoencoder with original vessels [ 17 ]. This method of vessel extracction in retinal images are used for classification. While deep learning models are highly chosen in retinal fundus image clasification, GAN are evolving in large areas due to its capacity in generating high resolution images from low quality images, generating augmented data and various applications [ 18 ]. The selection of appropriate model for retinal classification could save time and produce accurate results. The new multichannel multiple landmark is used for genrating fundus images from combination of vessels, optic disk that used pix2pix with MobileNet that achieved high performance that proves high resolution images are more beneficial for improving performance [ 19 , 20 ]. Another study used GAN for generating synthetic images which was trained on RetCam which are retinopahty of prematurity. This study suggest GAN generated images have potential applications. The SS-DCGAN classifier was employed for larger dataset that is highly efficient form retinal image classification. GAN employed image segmentation and augmentation for retinal classification are vastly studied. GAN are highly recommended for generating synthetic images form real data. Incase of glaucoma detection GANs are highly chosen for analysis [ 21 ]. The use of DCGAN for synthesizing MRI is highly focused on brain tumour which shows the wide applicability of GAN for medical sector [ 22 ]. GAN is highly suitbale for healthcare sector where the data generated is of limited nunmbers. Even GAN are proposed for generating CAPTCHA, in which the original image and generated images are compared [ 23 ]. The privacy concern ia also highly preserved while using GAN which is why it is highly recommended in healthcare sectors. This is mostly used in retinal fundus image classification. Machine learning models suffer due to less number of data which is rectified with GAN, that generates more realistic images from original data. While training with large and sufficient number of images the model becomes more robust and provide accurate results [ 25 ]. 3 Proposed Work This proposed work concentrates on retinal image classification with generative adversial network that augments the retinal images for more precise and accurate classification. Generative AI with GAN are applied to medical imaging like OCT, fundus photography, ultrasound with orbital images and MRI analysis [ 26 ]. This proposed work employs retinal image augmentation with the following approaches. Data Augmentation Techniques Deep Convolutional Generative Adversial Networks This is a model that has a convolutional decoder neural network for a generator. This is specially designed to optimize the performance of the unsupervised learning. This method capitalize the spatial upsampling in decoder operation in the generator, while this helps in generation of higher resolution images. Refining the model produces high quality images. The major modifications include strided convolutions in the discriminator. Another incorporation include batch normalization in generator and discriminator that helps in distribution of generated and actual images. This also eliminates connected hidden layers making it more convenient. ReLu activation is adopted in generator where Tanh is employed. In discriminator LeakyReLu is used. This is mostly used for producing high quality images [ 27 ]. Figure 2 depicts the overall architecture of the GAN with its major components a) Generator and b) Discriminator. The objective function of CGAN is elaborated below. Let us consider the loss function of discriminator as which show the variation between two functional parameters. L D = error(D(x),1) + error(D(G(z)),0) Eq. (1) The role of discriminator is to analyze whether the genrerated image is real or fake. It is a must to minimize the loss function. G must reduce thte discrepancy between label 1, actual image and generated image. Lg = error(D(G(z)),1) Eq. (2) With both the loss functions Eq. (1,2)it is easier to train the discriminator and generated images. The two loss functions have some difference than the original one which is represented as max D {log(D(x) + log(1-D(G(z)))} Eq. (3) In this case of Eq. (3), only loss function is given as opposed to the maximization problem. Hence min-max came into effect. min G max D {log(D(x)) + log(1-D(G(z)))} Eq. (4) The loss for both G and D is provided since the gradient of the function y = log(1x). Hence maximizing log(D(G(z)) or minimizing log(D(G(z)) will surely yield better results. min G max D V(D,G)=E x∼ p data(x) [log D(x|y)] + E z ∼p z (z) [log(1-D(G(z|y)))] Eq. (5) Where pdata(x) – real data proabability distribution P z(z) - random noise probability distribution E – mathematical expectation. y-conditional variable. Auxiliary Classifier Generative Adversial Networks Auxiliary model has an auxiliary structure that is more similar to InfoGAN. This framework has limited information to the class label. Here discriminator has a classifier responsible for categorizing samples with sub-classes. This makes it conveniant for improving training stability. The model training involved often used images for most classes when there is increased labels. The reason behind this presence of auxiliary classifier. This models are mostly preferred for data augmentation and image synthesis. The generator employs noise along the categorial label C sampled from overall distribution PC, that results in generation of sample represented as Xfake = G(c,z). With the help of discriminator there is high possibility of discrimination between actual and generated images. The concerns include authenticity and category labels. The motive is to increase the combined Loss. While training of generator, the objective is to increase difference between loss. The noise is represented by Z is automatically acquired and remain unchanged by the class label [ 28 ]. The proposed work flow depicted in Fig. 3 elaborates the retinal image enhancement with various GAN models. The images are preprocessed with noise removal and then augmented with GAN and then evaluated with the augmented data. DRIVE dataset is used for the analysis. Image Enhancement GAN are employed for image enhancement and reconstructing images especially in medical domain. With dual network architecture that has generator and discriminator, the prior is used to generate images while the latter is used for authenticating the generated images. With reconstruction the model can produce better resolution and highly accurate images [ 29 ]. Super Resolution GAN This model involves enhancement of real implementation of generated images that brings perceptual loss in addition to adversial loss. Mean square error is the loss function chosen for analysis that results in loss of high frequency details in recovered images. The perceptual loss is formulated by contrasting different features form CNN for both real and generated images. SRGAN is highly suitable for restoring image details that is obtained from low quality inputs and also helps in building high quality images [ 30 ]. This can also be employed for CT and MRI images that is of low quality images. This method is highly suitable for generating realistic data in the generated images that compete the original images. Image to Image Enhancement Cycle Generative Adversial Network CycleGAN handles problems where independent domains are involved. This utilizes the couple of complementary GAN that form a cyclic network topology. This creates a mapping in various X and Y domains. With two generators G and F, two discriminator Dy and Dx, the G takes an input image from X domain and attempts to transfer it to domain Y. While the process is going on the discriminator Dy verifies whether the generated image is authentic. The generator F helps in converting image from Y domain to X domain with discriminator Dx determine the authencity of generated images [ 31 ]. Figure 4 compares the image before and after cycleGAN. Deep Learning Models The Dense Net and Efficient Net are taken for classification. These networks are analyzed for retinal images with 64*64 that is resized from the original dataset. The purpose of this is to enhance better discrimination with input size as a feature. Previous works have taken 28*28 that results in lesser evaluation, while this 64*64 pixel present more real images. The CNN model is composed of following layers in which Conv2D performs spatial convolution on 2D input data using filters to extract features that are highly relevant in image processing. MaxPooling2D is the majorly used for down-sampling feature maps. It has sliding window over input feature lie and choose maximum value for each window. Flatten layer convert data into one dimensional array. This connects the output of one layer to another layer as input with different shapes. There is a transition from convolutional layer to FC layers [ 32 ]. Dropout helps in preventing over fitting by randomly deactivating neurons during training. This helps in over fitting. Dense layers are used to transform high level features extracted by convolutional and pooling layers. Figure 5 explains the low resolution image to high resolution retinal image with the feature extraction and enhancement methods. 4 Results & Discussion This model has learned from underlying patterns in the data. This work contributes an improved retinal fundus image super resolution based on GAN. The results prove that the model is more ideal as it works well for new data. The resolution of the images also have an impact on the classification of the model. Higher resolution images are used to provide more detailed features that enhances CNN performance. Figure 6 explains the accuracy of the proposed model. Figure 7 shows the loss during training and validation with epochs. From the results it is evident that the accuracy of the model is improved and loss of the model is greatly decreased which showcases that the proposed model is highly effective. 5 Conclusion This work concludes that the GAN are highly used for generating synthetic images from original data. This is highly suitable for medical domain where the data is less and suffer from efficient analysis. Generating synthetic images using GAN is highly recommended in healthcare sector. Low resolution images are enhanced with the GAN based techniques. Among all the models, DCGAN based methods highly suitable for retinal image enhancement. Declarations Fund Declaration This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Author Contribution S. Arun Kumar, M. Karuppasamy and Subburaj conceived of the presented idea. S. Arun Kumar, M. Karuppasamy developed the theory and performed the computations. Subburaj David Neels Ponkumar and M. Satheesh Kumar verified the analytical methods. Subburaj. encouraged S. Arun Kumar, M. Karuppasamy to investigate Deep Convolutional Generative Adversial Networks and supervised the findings of this work. All authors discussed the results and contributed to the final manuscript. References Roppelt, J. S., Kanbach, D. K., & Kraus, S. (2024). Artificial intelligence in healthcare institutions: A systematic literature review on influencing factors. Technology in society , 76 , 102443. Jindal, J. A., Lungren, M. P., & Shah, N. H. (2024). Ensuring useful adoption of generative artificial intelligence in healthcare. Journal of the American Medical Informatics Association , 31 (6), 1441–1444. Song, Y., Liu, Y., Lin, Z., Zhou, J., Li, D., Zhou, T., & Leung, M. F. (2024). Learning from AI-generated annotations for medical image segmentation. IEEE Transactions on Consumer Electronics . Tan, H., Guo, Z., Lin, Z., Chen, Y., Huang, D., Yuan, W., … Yan, J. (2024). General generative AI-based image augmentation method for robust rooftop PV segmentation. Applied Energy , 368 , 123554. Rao, V. S., Balakrishna, R., El-Ebiary, Y. B., Thapar, P., Saravanan, K. A., & Godla, S. R. (2024). Ai driven anomaly detection in network traffic using hybrid cnn-gan. Journal of Advances in Information Technology , 15 (7), 886–895. Koetzier, L. R., Wu, J., Mastrodicasa, D., Lutz, A., Chung, M., Koszek, W. A., … Willemink, M. J. (2024). Generating synthetic data for medical imaging. Radiology , 312 (3), e232471. Bharti, I., Chauhan, K., & Aggarwal, P. (2025). Generative AI: Next Frontier for Competitive Advantage. In Enhancing Communication and Decision-Making With AI (pp. 1–36). IGI Global. Carayannis, E. G., Dumitrescu, R., Falkowski, T., & Zota, N. R. (2024). EMPOWERING SMES-“HARNESSING THE POTENTIAL OF GEN AI FOR RESILIENCE AND COMPETITIVENESS”. IEEE Transactions on Engineering Management . Srinivasan, S. M., & Sharma, V. (2025). Applications of AI in cardiovascular disease detection—A review of the specific ways in which AI is being used to detect and diagnose cardiovascular diseases. AI in Disease Detection: Advancements and Applications , 123–146 Jenelle A Jindal, Matthew P Lungren, Nigam H Shah, Ensuring useful adoption of generative artificial intelligence in healthcare, Journal of the American Medical Informatics Association , Volume 31, Issue 6, June 2024, Pages 1441–1444, https://doi.org/10.1093/jamia/ocae043 Chen, Y., Yang, X. H., Wei, Z., Heidari, A. A., Zheng, N., Li, Z., … Guan, Q. (2022). Generative adversarial networks in medical image augmentation: a review. Computers in Biology and Medicine , 144 , 105382. Showrov, A. A., Aziz, M. T., Nabil, H. R., Jim, J. R., Kabir, M. M., Mridha, M. F., … Shin, J. (2024). Generative adversarial networks (GANs) in medical imaging: advancements, applications and challenges. IEEE Access . Sabnam, S., & Rajagopal, S. (2024). Application of generative adversarial networks in image, face reconstruction and medical imaging: challenges and the current progress. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization , 12 (1), 2330524. Mercaldo, F., Martinelli, F., & Santone, A. (2024). Deep convolutional generative adversarial networks in image-based android malware detection. Computers , 13 (6), 154. Zhang, Y., Xie, H., Zhuang, S., & Zhan, X. (2024). Image Processing and Optimization Using Deep Learning-Based Generative Adversarial Networks (GANs). Journal of Artificial Intelligence General science (JAIGS) ISSN: 3006–4023 , 5 (1), 50–62. Zhao, L., Chi, H., Zhong, T., & Jia, Y. (2024). Perception-oriented generative adversarial network for retinal fundus image super-resolution. Computers in Biology and Medicine , 168 , 107708. Jia, Y., Chen, G., & Chi, H. (2024). Retinal fundus image super-resolution based on generative adversarial network guided with vascular structure prior. Scientific Reports , 14 (1), 22786. Xie, X., Jiachu, D., Liu, C., Xie, M., Guo, J., Cai, K., … Zheng, C. (2024). Generating Synthesized Fluorescein Angiography Images From Color Fundus Images by Generative Adversarial Networks for Macular Edema Assessment. Translational Vision Science & Technology , 13 (9), 26–26. Liu, L., Hong, J., Wu, Y., Liu, S., Wang, K., Li, M., … Wu, X. (2024). Digital ray: enhancing cataractous fundus images using style transfer generative adversarial networks to improve retinopathy detection. British Journal of Ophthalmology , 108 (10), 1423–1429 Govindharaj, I., Santhakumar, D., Pugazharasi, K., Ravichandran, S., Prabhu, R. V., & Raja, J. (2025). Enhancing glaucoma diagnosis: Generative adversarial networks in synthesized imagery and classification with pretrained MobileNetV2. MethodsX , 14 , 103116. Di Giammarco, M., Santone, A., Cesarelli, M., Martinelli, F., & Mercaldo, F. (2024). Evaluating Deep Learning Resilience in Retinal Fundus Classification with Generative Adversarial Networks Generated Images. Electronics , 13 (13), 2631. Powroźnik, P., Skublewska-Paszkowska, M., Nowomiejska, K., Aristidou, A., Panayides, A., & Rejdak, R. (2025). Deep convolutional generative adversarial networks in retinitis pigmentosa disease images augmentation and detection. Advances in Science and Technology. Research Journal , 19 (2). Ferrara, E. (2024). GenAI against humanity: Nefarious applications of generative artificial intelligence and large language models. Journal of Computational Social Science , 7 (1), 549–569. Kar, M. K., & Nath, M. K. (2024). Efficient segmentation of vessels and disc simultaneously using multi-channel generative adversarial network. SN Computer Science , 5 (3), 288. Chaudhuri, R., & Deb, S. (2024). Precise lesion analysis to detect diabetic retinopathy using Generative Adversarial Network (GAN) and Mask-RCNN. Procedia Computer Science , 235 , 520–529. Gencer, K., Gencer, G., Ceran, T. H., Bilir, A. E., & Doğan, M. (2025). PHOTODIAGNOSIS WITH DEEP LEARNING: A GAN AND AUTOENCODER-BASED APPROACH FOR DIABETIC RETINOPATHY DETECTION. Photodiagnosis and Photodynamic Therapy , 104552. Shoaib, M. R., Emara, H. M., Mubarak, A. S., Omer, O. A., Abd El-Samie, F. E., & Esmaiel, H. (2025). Revolutionizing diabetic retinopathy diagnosis through advanced deep learning techniques: Harnessing the power of GAN model with transfer learning and the DiaGAN-CNN model. Biomedical Signal Processing and Control , 99 , 106790. Showrov, A. A., Aziz, M. T., Nabil, H. R., Jim, J. R., Kabir, M. M., Mridha, M. F.,… Shin, J. (2024). Generative adversarial networks (GANs) in medical imaging: advancements,applications and challenges. IEEE Access . Liu, L., Jan, H., Tang, C., He, H., Zhang, L., & Lei, Z. (2024). Dual-channel lightweight GAN for enhancing color retinal images with noise suppression and structural protection. Journal of the Optical Society of America A , 41 (10), 1948–1958. Chaudhuri, R., & Deb, S. (2024). Precise lesion analysis to detect diabetic retinopathy using Generative Adversarial Network (GAN) and Mask-RCNN. Procedia Computer Science , 235 , 520–529. Kim, Y. E., Lee, E. J., Yoon, J. S., Kwak, J., & Kim, H. (2025). Optical Coherence Tomography Image Enhancement and Layer Detection Using Cycle-GAN. Diagnostics , 15 (3), 277. Shoaib, M. R., Emara, H. M., Mubarak, A. S., Omer, O. A., Abd El-Samie, F. E., & Esmaiel, H. (2025). Revolutionizing diabetic retinopathy diagnosis through advanced deep learning techniques: Harnessing the power of GAN model with transfer learning and the DiaGAN-CNN model. Biomedical Signal Processing and Control , 99 , 106790. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7017188","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":500110702,"identity":"82b9c088-f29f-4799-8f61-9446433140cc","order_by":0,"name":"S. Arun Kumar","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYPCCA4z9ICqhgBQtMxtAWgxI0bLhAIgmRgv/7DNmD3/uuCO7+fzqxA8PDBjk+cUO4NcicS7H3Jj3zDPjbTfebpYAOsxw5uwEAtac4TGTZmw7nLjtxtkNIC0JBrcJaJEHapH8CdSyecbZzT+I0mIA1CLBC9Sygb93G3G2GJ5hK5PmPXPYeMYN3m0WCQYShP0id4Z5m+TPHYdl+/vPbr75o8JGnl+agBYwYGwAEhJglRJEKIdr4T9ApOpRMApGwSgYcQAAqLJJjha0mZEAAAAASUVORK5CYII=","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"S.","middleName":"Arun","lastName":"Kumar","suffix":""},{"id":500110703,"identity":"2492ed4d-3b52-4362-b88f-20a9012be8b5","order_by":1,"name":"M. Karuppasamy","email":"","orcid":"","institution":"ACS College of Engineering","correspondingAuthor":false,"prefix":"","firstName":"M.","middleName":"","lastName":"Karuppasamy","suffix":""},{"id":500110704,"identity":"0f704db9-4f4c-485e-95f7-fa638f3bea74","order_by":2,"name":"T. Subburaj","email":"","orcid":"","institution":"Rajarajeswari college of engineering","correspondingAuthor":false,"prefix":"","firstName":"T.","middleName":"","lastName":"Subburaj","suffix":""},{"id":500110706,"identity":"d40ea395-73e0-48a2-ab33-b0d03c533b18","order_by":3,"name":"David Neels Ponkumar","email":"","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"Neels","lastName":"Ponkumar","suffix":""},{"id":500110708,"identity":"76b3c90c-4247-4dcf-abdd-94ea26e08c93","order_by":4,"name":"M. Satheesh Kumar","email":"","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"M.","middleName":"Satheesh","lastName":"Kumar","suffix":""},{"id":500110711,"identity":"85909c76-62e4-496a-abcc-c995437c4171","order_by":5,"name":"M Jansi Rani","email":"","orcid":"","institution":"Mepco Schlenk Engineering College","correspondingAuthor":false,"prefix":"","firstName":"M","middleName":"Jansi","lastName":"Rani","suffix":""}],"badges":[],"createdAt":"2025-07-01 07:23:33","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7017188/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7017188/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":89273445,"identity":"4f44a606-a87c-45cf-a2d7-ffec4f79c17a","added_by":"auto","created_at":"2025-08-18 09:09:12","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":63340,"visible":true,"origin":"","legend":"\u003cp\u003eApplications of GAN in Healthcare Domain\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/85efe7e8bf42fa71311e79bd.jpg"},{"id":89273504,"identity":"83fb948d-3731-4ca1-8f59-48e329cc3690","added_by":"auto","created_at":"2025-08-18 09:09:18","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":160600,"visible":true,"origin":"","legend":"\u003cp\u003eDemonstration of Image generation using GAN\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/119ffc6b5ad7330d8b0c7045.jpg"},{"id":89273457,"identity":"f36009ed-a2af-4931-9085-dbae7ec34d76","added_by":"auto","created_at":"2025-08-18 09:09:14","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":49687,"visible":true,"origin":"","legend":"\u003cp\u003eProposed GAN based Retinal Detection\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/08e62b844ab18fe98bc56a7a.jpg"},{"id":89273486,"identity":"899d2bb1-13b0-432f-9d09-c653119fbcda","added_by":"auto","created_at":"2025-08-18 09:09:16","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":146306,"visible":true,"origin":"","legend":"\u003cp\u003eLow Resolution Image after Enhancement with GAN\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/05cfa738cbcbece860416137.jpg"},{"id":89273444,"identity":"be209f05-412a-41c1-946b-0da13520474d","added_by":"auto","created_at":"2025-08-18 09:09:12","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":223647,"visible":true,"origin":"","legend":"\u003cp\u003eImage Enhanced with Deep Learning Model\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/d53df1e68ae7feafb2423fa4.jpg"},{"id":89273450,"identity":"e00b8ea6-f1b7-41a4-ab9e-b6f15ec87669","added_by":"auto","created_at":"2025-08-18 09:09:13","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":43068,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and Validation Accuracy\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/3b11f20b4fbf901d15169d6c.jpg"},{"id":89273496,"identity":"5967198d-b973-43fc-a505-07307adef7b6","added_by":"auto","created_at":"2025-08-18 09:09:17","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":36480,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and Validation loss\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/a467f662a82c86183515f7c7.jpg"},{"id":92355218,"identity":"7a738712-27aa-45d4-ba22-bd8c6949d5c4","added_by":"auto","created_at":"2025-09-28 14:01:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1266826,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7017188/v1/e2889bf3-4b45-4450-b9b2-ea3be8146827.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"GAN-DCNN: A Generative Adversial Network based Deep Convolutional Neural Network for Retinal Vessel Enhancement","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eArtificial intelligence has become the most popular tool in various application areas which include healthcare sector that helps in clinical decision making and improving health outcomes [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Generative AI is one of the AI area which generates new data and images that assist patient care, providing clinical decision making [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn the present era, Generative Adversial Networks have evolved as a vital tool in the medical imaging and healthcare data analytics. It is used in variety of applications like medical image synthesis, data augmentation, anamoly detection and disease diagnosis [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u0026amp; \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. The indepth training mechanism empoers GAN to observe intricate data distribution and produce quality that helps in training data. With the use of various medical images, GAN generates synthetic images with diverse variance and also helps in facilitating the augmentation of training data. With advanced deep learning models GAN also helps in better prediction if diseses especially in medical imaging. GAN generates realistic synthetic images that include pathological images which contributes to the study of rare diseases and abnormalities [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. GAN plays a diverse role in medical image enhancement and reconstruction [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. GAN methodologies convert nosiy images into high quality images that enhances the procedures and treatment strategies. GAN are used in various health related issues like cardiovascular diseases, diabetes, cancer and medical image analysis [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the application areas of GAN in medical and healthcare sectors. GAN uses the training of two neural networks for image generation and discrimination. The generative model uses training data to produce images while discriminative model estimates the image came from training data.\u003c/p\u003e\u003cp\u003e\u003cb\u003eAugmentation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAugmentation is majorly used in data and image processing that helps in medical image analysis where there is limited dataset [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. This involves various features like rotation, scaling and flipping to elaborate the dataset size. Incase of medical imaging there will be shortage of images to train a mode where there will be less accurate results, while augmentation helps in overcoming this issue. Augmentation expands the dataset size required for efficient training and provide more robust and accurate results [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cb\u003eSegmentation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eSegmentation in medical images is used to improve the image quality. This also allows multi-modal fusion, helps in image reconstruction and enhance the medical image analysis. Medical diagnoses are more perfect with reliable results when segmentation techniques are involved [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cb\u003eReconstruction\u003c/b\u003e\u003c/p\u003e\u003cp\u003eIt is a process that helps in reconstructing an image with the help of algorithm that builds image from an incomplete data. This is mostly used in medical imaging, remote sensing and computer vision. This process usually generated high-fidelity images from low resolution images [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. This also removes noise using generative adversial networks. GAN consists of generator and discriminator. Generator learns to real images from noisy or imperfect images. In particular with medical image analysis, GAN generated images are valuable like super-resolution images with enhanced image quality. This assist the medical sector in analyzing medical scans with clear and detailed images for accurate results.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDetection\u003c/b\u003e\u003c/p\u003e\u003cp\u003eGAN has become a vital tool in medical image analysis. This also helps in detection of diseases or medical conditions such as cancer, retinal abnormalities, Alzhemier\u0026rsquo;s and other anomalies. GAN generates synthetic images of real data and compares with the real medical images. This enhances the accuracy of early disease detection [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. If diagnosis is done earlier there is high possibility of reduction of risk in major diseases.\u003c/p\u003e\u003cp\u003e\u003cb\u003eClassification\u003c/b\u003e\u003c/p\u003e\u003cp\u003eGAN is an indirect mechanism for classification by diversifying training data through data augmentation. Though GAN does not directly support classification, it is a hidden factor that enhances feature extraction, detect anomalies and also provide test samples in case of semi-supervised learning. With these process the classification accuracy is much improved and utilized for various applications. Deep learning based medical image classification is also improved with GAN [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cb\u003eImage Synthesis\u003c/b\u003e\u003c/p\u003e\u003cp\u003eGAN is used to generate synthetic medical images that represents similar characteristics of patient data. It contains generator that generates images and discriminator that differentiate between real and synthetic images. In healthcare sectors, medical image synthesis enables training of the machine learning models. With the help of data or image generation the size of the dataset is also increased making it easier for training machine learning models, at the same time the privacy of the data is also conserved.\u003c/p\u003e\u003cp\u003eGAI is specially more helpful for personalized treatment that involves predictive modeling to featuring treatment plans. Conditional Variant Encoder combines conditional variables and autoencoders to learn patient information that incorporated specific conditions for treatment plans. By specifying different conditions the model could generate various treatment strategies to individual patients. Fine Tuned NLP such as BERT are improved in generating personlaized plans for patient information. GAI with GAN are almost used for medical image analysis. It is also helpful in monitoring various health conditions, diabetes analysis, medical image analysis and cancer detection.\u003c/p\u003e"},{"header":"2 Related Work","content":"\u003cp\u003eThere were several applications of GAN in retinal image classification that include segmentation, data augmentation, producing high resolution images, prediction and feature extraction. Perception oriented GAN for retinal images that provide high resolution images are employed [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Another study employed adversial technique on vascular network with some concerns overcome by training autoencoder with original vessels [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. This method of vessel extracction in retinal images are used for classification.\u003c/p\u003e\u003cp\u003eWhile deep learning models are highly chosen in retinal fundus image clasification, GAN are evolving in large areas due to its capacity in generating high resolution images from low quality images, generating augmented data and various applications [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The selection of appropriate model for retinal classification could save time and produce accurate results. The new multichannel multiple landmark is used for genrating fundus images from combination of vessels, optic disk that used pix2pix with MobileNet that achieved high performance that proves high resolution images are more beneficial for improving performance [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Another study used GAN for generating synthetic images which was trained on RetCam which are retinopahty of prematurity. This study suggest GAN generated images have potential applications.\u003c/p\u003e\u003cp\u003eThe SS-DCGAN classifier was employed for larger dataset that is highly efficient form retinal image classification. GAN employed image segmentation and augmentation for retinal classification are vastly studied. GAN are highly recommended for generating synthetic images form real data. Incase of glaucoma detection GANs are highly chosen for analysis [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe use of DCGAN for synthesizing MRI is highly focused on brain tumour which shows the wide applicability of GAN for medical sector [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. GAN is highly suitbale for healthcare sector where the data generated is of limited nunmbers. Even GAN are proposed for generating CAPTCHA, in which the original image and generated images are compared [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. The privacy concern ia also highly preserved while using GAN which is why it is highly recommended in healthcare sectors. This is mostly used in retinal fundus image classification. Machine learning models suffer due to less number of data which is rectified with GAN, that generates more realistic images from original data. While training with large and sufficient number of images the model becomes more robust and provide accurate results [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e"},{"header":"3 Proposed Work","content":"\u003cp\u003eThis proposed work concentrates on retinal image classification with generative adversial network that augments the retinal images for more precise and accurate classification.\u003c/p\u003e\u003cp\u003eGenerative AI with GAN are applied to medical imaging like OCT, fundus photography, ultrasound with orbital images and MRI analysis [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis proposed work employs retinal image augmentation with the following approaches.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData Augmentation Techniques\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eDeep Convolutional Generative Adversial Networks\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis is a model that has a convolutional decoder neural network for a generator. This is specially designed to optimize the performance of the unsupervised learning. This method capitalize the spatial upsampling in decoder operation in the generator, while this helps in generation of higher resolution images. Refining the model produces high quality images. The major modifications include strided convolutions in the discriminator. Another incorporation include batch normalization in generator and discriminator that helps in distribution of generated and actual images. This also eliminates connected hidden layers making it more convenient. ReLu activation is adopted in generator where Tanh is employed. In discriminator LeakyReLu is used. This is mostly used for producing high quality images [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e depicts the overall architecture of the GAN with its major components a) Generator and b) Discriminator. The objective function of CGAN is elaborated below.\u003c/p\u003e\u003cp\u003eLet us consider the loss function of discriminator as which show the variation between two functional parameters.\u003c/p\u003e\u003cp\u003eL\u003csub\u003eD\u003c/sub\u003e= error(D(x),1)\u0026thinsp;+\u0026thinsp;error(D(G(z)),0) \u003cb\u003eEq.\u0026nbsp;(1)\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe role of discriminator is to analyze whether the genrerated image is real or fake. It is a must to minimize the loss function. G must reduce thte discrepancy between label 1, actual image and generated image.\u003c/p\u003e\u003cp\u003eLg\u0026thinsp;=\u0026thinsp;error(D(G(z)),1) \u003cb\u003eEq.\u0026nbsp;(2)\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWith both the loss functions Eq.\u0026nbsp;(1,2)it is easier to train the discriminator and generated images.\u003c/p\u003e\u003cp\u003eThe two loss functions have some difference than the original one which is represented as\u003c/p\u003e\u003cp\u003emax\u003csub\u003eD\u003c/sub\u003e {log(D(x)\u0026thinsp;+\u0026thinsp;log(1-D(G(z)))} \u003cb\u003eEq.\u0026nbsp;(3)\u003c/b\u003e\u003c/p\u003e\u003cp\u003eIn this case of Eq.\u0026nbsp;(3), only loss function is given as opposed to the maximization problem. Hence min-max came into effect.\u003c/p\u003e\u003cp\u003emin\u003csub\u003eG\u003c/sub\u003emax\u003csub\u003eD\u003c/sub\u003e{log(D(x))\u0026thinsp;+\u0026thinsp;log(1-D(G(z)))} \u003cb\u003eEq.\u0026nbsp;(4)\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe loss for both G and D is provided since the gradient of the function y\u0026thinsp;=\u0026thinsp;log(1x). Hence maximizing log(D(G(z)) or minimizing log(D(G(z)) will surely yield better results.\u003c/p\u003e\u003cp\u003emin\u003csub\u003eG\u003c/sub\u003e max\u003csub\u003eD V(D,G)=E x\u0026sim;\u003c/sub\u003ep\u003csub\u003edata(x)\u003c/sub\u003e [log D(x|y)]\u0026thinsp;+\u0026thinsp;E\u003csub\u003ez\u003c/sub\u003e\u0026sim;p z (z) [log(1-D(G(z|y)))] \u003cb\u003eEq.\u0026nbsp;(5)\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWhere pdata(x) \u0026ndash; real data proabability distribution\u003c/p\u003e\u003cp\u003eP z(z) - random noise probability distribution\u003c/p\u003e\u003cp\u003eE \u0026ndash; mathematical expectation.\u003c/p\u003e\u003cp\u003ey-conditional variable.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eAuxiliary Classifier Generative Adversial Networks\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAuxiliary model has an auxiliary structure that is more similar to InfoGAN. This framework has limited information to the class label. Here discriminator has a classifier responsible for categorizing samples with sub-classes. This makes it conveniant for improving training stability. The model training involved often used images for most classes when there is increased labels. The reason behind this presence of auxiliary classifier. This models are mostly preferred for data augmentation and image synthesis. The generator employs noise along the categorial label C sampled from overall distribution PC, that results in generation of sample represented as Xfake\u0026thinsp;=\u0026thinsp;G(c,z). With the help of discriminator there is high possibility of discrimination between actual and generated images. The concerns include authenticity and category labels. The motive is to increase the combined Loss. While training of generator, the objective is to increase difference between loss. The noise is represented by Z is automatically acquired and remain unchanged by the class label [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe proposed work flow depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e elaborates the retinal image enhancement with various GAN models. The images are preprocessed with noise removal and then augmented with GAN and then evaluated with the augmented data. DRIVE dataset is used for the analysis.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eImage Enhancement\u003c/b\u003e\u003c/p\u003e\u003cp\u003eGAN are employed for image enhancement and reconstructing images especially in medical domain. With dual network architecture that has generator and discriminator, the prior is used to generate images while the latter is used for authenticating the generated images. With reconstruction the model can produce better resolution and highly accurate images [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cb\u003eSuper Resolution GAN\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis model involves enhancement of real implementation of generated images that brings perceptual loss in addition to adversial loss. Mean square error is the loss function chosen for analysis that results in loss of high frequency details in recovered images. The perceptual loss is formulated by contrasting different features form CNN for both real and generated images. SRGAN is highly suitable for restoring image details that is obtained from low quality inputs and also helps in building high quality images [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. This can also be employed for CT and MRI images that is of low quality images. This method is highly suitable for generating realistic data in the generated images that compete the original images.\u003c/p\u003e\u003cp\u003e\u003cb\u003eImage to Image Enhancement\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eCycle Generative Adversial Network\u003c/b\u003e\u003c/p\u003e\u003cp\u003eCycleGAN handles problems where independent domains are involved. This utilizes the couple of complementary GAN that form a cyclic network topology. This creates a mapping in various X and Y domains. With two generators G and F, two discriminator Dy and Dx, the G takes an input image from X domain and attempts to transfer it to domain Y. While the process is going on the discriminator Dy verifies whether the generated image is authentic. The generator F helps in converting image from Y domain to X domain with discriminator Dx determine the authencity of generated images [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e compares the image before and after cycleGAN.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eDeep Learning Models\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe Dense Net and Efficient Net are taken for classification. These networks are analyzed for retinal images with 64*64 that is resized from the original dataset. The purpose of this is to enhance better discrimination with input size as a feature. Previous works have taken 28*28 that results in lesser evaluation, while this 64*64 pixel present more real images. The CNN model is composed of following layers in which Conv2D performs spatial convolution on 2D input data using filters to extract features that are highly relevant in image processing. MaxPooling2D is the majorly used for down-sampling feature maps. It has sliding window over input feature lie and choose maximum value for each window. Flatten layer convert data into one dimensional array. This connects the output of one layer to another layer as input with different shapes. There is a transition from convolutional layer to FC layers [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. Dropout helps in preventing over fitting by randomly deactivating neurons during training. This helps in over fitting. Dense layers are used to transform high level features extracted by convolutional and pooling layers. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e explains the low resolution image to high resolution retinal image with the feature extraction and enhancement methods.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"4 Results \u0026 Discussion","content":"\u003cp\u003eThis model has learned from underlying patterns in the data. This work contributes an improved retinal fundus image super resolution based on GAN. The results prove that the model is more ideal as it works well for new data. The resolution of the images also have an impact on the classification of the model. Higher resolution images are used to provide more detailed features that enhances CNN performance. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e explains the accuracy of the proposed model.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows the loss during training and validation with epochs. From the results it is evident that the accuracy of the model is improved and loss of the model is greatly decreased which showcases that the proposed model is highly effective.\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eThis work concludes that the GAN are highly used for generating synthetic images from original data. This is highly suitable for medical domain where the data is less and suffer from efficient analysis. Generating synthetic images using GAN is highly recommended in healthcare sector. Low resolution images are enhanced with the GAN based techniques. Among all the models, DCGAN based methods highly suitable for retinal image enhancement.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFund Declaration\u003c/h2\u003e\n\u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eS. Arun Kumar, M. Karuppasamy and Subburaj conceived of the presented idea. S. Arun Kumar, M. Karuppasamy developed the theory and performed the computations. Subburaj David Neels Ponkumar and M. Satheesh Kumar verified the analytical methods. Subburaj. encouraged S. Arun Kumar, M. Karuppasamy to investigate Deep Convolutional Generative Adversial Networks and supervised the findings of this work. All authors discussed the results and contributed to the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eRoppelt, J. S., Kanbach, D. K., \u0026amp; Kraus, S. (2024). Artificial intelligence in healthcare institutions: A systematic literature review on influencing factors. \u003cem\u003eTechnology in society\u003c/em\u003e, \u003cem\u003e76\u003c/em\u003e, 102443.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJindal, J. A., Lungren, M. P., \u0026amp; Shah, N. H. (2024). Ensuring useful adoption of generative artificial intelligence in healthcare. \u003cem\u003eJournal of the American Medical Informatics Association\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(6), 1441\u0026ndash;1444.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSong, Y., Liu, Y., Lin, Z., Zhou, J., Li, D., Zhou, T., \u0026amp; Leung, M. F. (2024). Learning from AI-generated annotations for medical image segmentation. \u003cem\u003eIEEE Transactions on Consumer Electronics\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTan, H., Guo, Z., Lin, Z., Chen, Y., Huang, D., Yuan, W., \u0026hellip; Yan, J. (2024). General generative AI-based image augmentation method for robust rooftop PV segmentation. \u003cem\u003eApplied Energy\u003c/em\u003e, \u003cem\u003e368\u003c/em\u003e, 123554.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRao, V. S., Balakrishna, R., El-Ebiary, Y. B., Thapar, P., Saravanan, K. A., \u0026amp; Godla, S. R. (2024). Ai driven anomaly detection in network traffic using hybrid cnn-gan. \u003cem\u003eJournal of Advances in Information Technology\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(7), 886\u0026ndash;895.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKoetzier, L. R., Wu, J., Mastrodicasa, D., Lutz, A., Chung, M., Koszek, W. A., \u0026hellip; Willemink, M. J. (2024). Generating synthetic data for medical imaging. \u003cem\u003eRadiology\u003c/em\u003e, \u003cem\u003e312\u003c/em\u003e(3), e232471.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBharti, I., Chauhan, K., \u0026amp; Aggarwal, P. (2025). Generative AI: Next Frontier for Competitive Advantage. In \u003cem\u003eEnhancing Communication and Decision-Making With AI\u003c/em\u003e (pp. 1\u0026ndash;36). IGI Global.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCarayannis, E. G., Dumitrescu, R., Falkowski, T., \u0026amp; Zota, N. R. (2024). EMPOWERING SMES-\u0026ldquo;HARNESSING THE POTENTIAL OF GEN AI FOR RESILIENCE AND COMPETITIVENESS\u0026rdquo;. \u003cem\u003eIEEE Transactions on Engineering Management\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSrinivasan, S. M., \u0026amp; Sharma, V. (2025). Applications of AI in cardiovascular disease detection\u0026mdash;A review of the specific ways in which AI is being used to detect and diagnose cardiovascular diseases. \u003cem\u003eAI in Disease Detection: Advancements and Applications\u003c/em\u003e, 123\u0026ndash;146\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJenelle A Jindal, Matthew P Lungren, Nigam H Shah, Ensuring useful adoption of generative artificial intelligence in healthcare, \u003cem\u003eJournal of the American Medical Informatics Association\u003c/em\u003e, Volume 31, Issue 6, June 2024, Pages 1441\u0026ndash;1444, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/jamia/ocae043\u003c/span\u003e\u003cspan address=\"10.1093/jamia/ocae043\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, Y., Yang, X. H., Wei, Z., Heidari, A. A., Zheng, N., Li, Z., \u0026hellip; Guan, Q. (2022). Generative adversarial networks in medical image augmentation: a review. \u003cem\u003eComputers in Biology and Medicine\u003c/em\u003e, \u003cem\u003e144\u003c/em\u003e, 105382.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShowrov, A. A., Aziz, M. T., Nabil, H. R., Jim, J. R., Kabir, M. M., Mridha, M. F., \u0026hellip; Shin, J. (2024). Generative adversarial networks (GANs) in medical imaging: advancements, applications and challenges. \u003cem\u003eIEEE Access\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSabnam, S., \u0026amp; Rajagopal, S. (2024). Application of generative adversarial networks in image, face reconstruction and medical imaging: challenges and the current progress. \u003cem\u003eComputer Methods in Biomechanics and Biomedical Engineering: Imaging \u0026amp; Visualization\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(1), 2330524.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMercaldo, F., Martinelli, F., \u0026amp; Santone, A. (2024). Deep convolutional generative adversarial networks in image-based android malware detection. \u003cem\u003eComputers\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(6), 154.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, Y., Xie, H., Zhuang, S., \u0026amp; Zhan, X. (2024). Image Processing and Optimization Using Deep Learning-Based Generative Adversarial Networks (GANs). \u003cem\u003eJournal of Artificial Intelligence General science (JAIGS) ISSN: 3006\u0026ndash;4023\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(1), 50\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao, L., Chi, H., Zhong, T., \u0026amp; Jia, Y. (2024). Perception-oriented generative adversarial network for retinal fundus image super-resolution. \u003cem\u003eComputers in Biology and Medicine\u003c/em\u003e, \u003cem\u003e168\u003c/em\u003e, 107708.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJia, Y., Chen, G., \u0026amp; Chi, H. (2024). Retinal fundus image super-resolution based on generative adversarial network guided with vascular structure prior. \u003cem\u003eScientific Reports\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(1), 22786.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXie, X., Jiachu, D., Liu, C., Xie, M., Guo, J., Cai, K., \u0026hellip; Zheng, C. (2024). Generating Synthesized Fluorescein Angiography Images From Color Fundus Images by Generative Adversarial Networks for Macular Edema Assessment. \u003cem\u003eTranslational Vision Science \u0026amp; Technology\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(9), 26\u0026ndash;26.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, L., Hong, J., Wu, Y., Liu, S., Wang, K., Li, M., \u0026hellip; Wu, X. (2024). Digital ray: enhancing cataractous fundus images using style transfer generative adversarial networks to improve retinopathy detection. \u003cem\u003eBritish Journal of Ophthalmology\u003c/em\u003e, \u003cem\u003e108\u003c/em\u003e(10), 1423\u0026ndash;1429\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGovindharaj, I., Santhakumar, D., Pugazharasi, K., Ravichandran, S., Prabhu, R. V., \u0026amp; Raja, J. (2025). Enhancing glaucoma diagnosis: Generative adversarial networks in synthesized imagery and classification with pretrained MobileNetV2. \u003cem\u003eMethodsX\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e, 103116.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDi Giammarco, M., Santone, A., Cesarelli, M., Martinelli, F., \u0026amp; Mercaldo, F. (2024). Evaluating Deep Learning Resilience in Retinal Fundus Classification with Generative Adversarial Networks Generated Images. \u003cem\u003eElectronics\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(13), 2631.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePowroźnik, P., Skublewska-Paszkowska, M., Nowomiejska, K., Aristidou, A., Panayides, A., \u0026amp; Rejdak, R. (2025). Deep convolutional generative adversarial networks in retinitis pigmentosa disease images augmentation and detection. \u003cem\u003eAdvances in Science and Technology. Research Journal\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(2).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFerrara, E. (2024). GenAI against humanity: Nefarious applications of generative artificial intelligence and large language models. \u003cem\u003eJournal of Computational Social Science\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(1), 549\u0026ndash;569.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKar, M. K., \u0026amp; Nath, M. K. (2024). Efficient segmentation of vessels and disc simultaneously using multi-channel generative adversarial network. \u003cem\u003eSN Computer Science\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(3), 288.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChaudhuri, R., \u0026amp; Deb, S. (2024). Precise lesion analysis to detect diabetic retinopathy using Generative Adversarial Network (GAN) and Mask-RCNN. \u003cem\u003eProcedia Computer Science\u003c/em\u003e, \u003cem\u003e235\u003c/em\u003e, 520\u0026ndash;529.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGencer, K., Gencer, G., Ceran, T. H., Bilir, A. E., \u0026amp; Doğan, M. (2025). PHOTODIAGNOSIS WITH DEEP LEARNING: A GAN AND AUTOENCODER-BASED APPROACH FOR DIABETIC RETINOPATHY DETECTION. \u003cem\u003ePhotodiagnosis and Photodynamic Therapy\u003c/em\u003e, 104552.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShoaib, M. R., Emara, H. M., Mubarak, A. S., Omer, O. A., Abd El-Samie, F. E., \u0026amp; Esmaiel, H. (2025). Revolutionizing diabetic retinopathy diagnosis through advanced deep learning techniques: Harnessing the power of GAN model with transfer learning and the DiaGAN-CNN model. \u003cem\u003eBiomedical Signal Processing and Control\u003c/em\u003e, \u003cem\u003e99\u003c/em\u003e, 106790.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShowrov, A. A., Aziz, M. T., Nabil, H. R., Jim, J. R., Kabir, M. M., Mridha, M. F.,\u0026hellip; Shin, J. (2024). Generative adversarial networks (GANs) in medical imaging: advancements,applications and challenges. \u003cem\u003eIEEE Access\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, L., Jan, H., Tang, C., He, H., Zhang, L., \u0026amp; Lei, Z. (2024). Dual-channel lightweight GAN for enhancing color retinal images with noise suppression and structural protection. \u003cem\u003eJournal of the Optical Society of America A\u003c/em\u003e, \u003cem\u003e41\u003c/em\u003e(10), 1948\u0026ndash;1958.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChaudhuri, R., \u0026amp; Deb, S. (2024). Precise lesion analysis to detect diabetic retinopathy using Generative Adversarial Network (GAN) and Mask-RCNN. \u003cem\u003eProcedia Computer Science\u003c/em\u003e, \u003cem\u003e235\u003c/em\u003e, 520\u0026ndash;529.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, Y. E., Lee, E. J., Yoon, J. S., Kwak, J., \u0026amp; Kim, H. (2025). Optical Coherence Tomography Image Enhancement and Layer Detection Using Cycle-GAN. \u003cem\u003eDiagnostics\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(3), 277.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShoaib, M. R., Emara, H. M., Mubarak, A. S., Omer, O. A., Abd El-Samie, F. E., \u0026amp; Esmaiel, H. (2025). Revolutionizing diabetic retinopathy diagnosis through advanced deep learning techniques: Harnessing the power of GAN model with transfer learning and the DiaGAN-CNN model. \u003cem\u003eBiomedical Signal Processing and Control\u003c/em\u003e, \u003cem\u003e99\u003c/em\u003e, 106790.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Generative AI, generative adversial network, healthcare, retinal images, disease prediction, DCGAN, SRGAN","lastPublishedDoi":"10.21203/rs.3.rs-7017188/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7017188/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eArtificial intelligence is taking over the technology transmission in all areas. Generative AI is now mostly used in the healthcare sector due to its vivid application area. The emergence of Generative Adversial Network (GAN) in medical domain is highly recommended. The proposed work brings the essence of deep convolutional network that differentiates real and generated retinal images. This work highlights the advancement of GAN over other artificial intelligence models. By identifying appropriate networks of GAN, retinal images are generated in high quality. Low resolution images are enhanced with GAN models. The proposed work involves basic preprocessing, augmentation and enhancement with the retinal images that provide better accuracy with epochs. Deep learning models with GAN implementation are highly preferred for the retinal classification.\u003c/p\u003e","manuscriptTitle":"GAN-DCNN: A Generative Adversial Network based Deep Convolutional Neural Network for Retinal Vessel Enhancement","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-18 09:08:54","doi":"10.21203/rs.3.rs-7017188/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"921d0abb-d8cd-4950-9873-19d8ad55dc27","owner":[],"postedDate":"August 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-28T13:53:34+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-18 09:08:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7017188","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7017188","identity":"rs-7017188","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}