Publication: Modified residual block fully convolutional network for localization of neovascularization lesions from fundus images
Date
2023-10-01
Authors
Michael Tang Chi Seng
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Proliferative Diabetic Retinopathy (PDR) is a severe retinal disease that can lead to blindness among diabetic patients. Neovascularization is a characteristic feature of PDR. Its clinical manifestations include a high degree of retinal neovascularization and fibrous spreads. Currently used PDR diagnostic strategies rely on an ophthalmologist’s interpretation of a patient’s fundus image. However, this manual diagnosis requires effort. As a result, automated PDR detection is integrated into computer-aided diagnostic systems to help the ophthalmologists. Numerous deep learning techniques have been proposed for using fundus images to detect and diagnose neovascularization. Furthermore, the methods cannot segment the pixels of neovascularization lesions precisely. In addition, some researchers employed transfer learning to train a pre-trained model to detect neovascularization. However, since pre-trained models were trained on a dataset containing objects dissimilar to those in the target dataset, transfer learning is ineffective. Therefore, this research aims to overcome the ineffectiveness of the transfer learning approach. Moreover, this research also aims to introduce effective image processing and deep learning techniques for detecting and localizing neovascularization. A novel, fully convolutional network is formulated to automatically detect and locate neovascularization lesions in fundus images, which is not possible with previously published convolution neural network models. The novelty of the fully convolutional network lies in the implementation of modified residual blocks proposed in this study, which distinguishes it from prior approaches and contributes to its originality. The model is composed of thirteen trainable convolutional layers. Overall, the contribution of this study is the formulation of a novel, fully convolutional network that outperforms previously used deep learning methods for detecting neovascularization. Second, a novel transfer learning method is proposed that avoids the need to train a new network from scratch. Third, a modified residual block is proposed to allow for downsampling, which produces better results than the original residual block. The proposed method achieves an average accuracy, sensitivity, specificity, precision, Jaccard similarity, and Dice similarity of 0.9877, 0.8057, 0.9903, 0.7424, 0.5830, and 0.6896, respectively. It outperformed existing models used for detecting neovascularization in terms of accuracy, sensitivity, and specificity. It can also segment the neovascularization regions, which is not possible with the existing methods.