Convolution neural network for diabetic retinopathy classification with select enhancment algorithm

Abstract

Diabetic retinopathy and retinal vascular occlusion are the most significant causes of vision loss. Physical examinations are no longer sufficient to detect early retinal diseases due to the rise in patients with diabetes and high blood pressure, making a multiclass automated detection system a necessity in the health care field. However, the majority of the previous proposed automated systems were designed to diagnose a specific case, single class, such as diabetic retinopathy or normal retina (no-diabetic retinopathy), where these previous automated systems require several complexes preprocessing steps, such as cropping the area of interest, which requires an expertise of ophthalmologist. Furthermore, the cropping method may result in the loss of important pieces of information about retina disease in the discarded area, so it is being a source of human error. Additionally, the new multiclass detection technique demands a high number of retina samples, which are tough to obtain in health care centres, in addition to mostly complex segmentation phases that should be done. In this thesis, a novel preprocessing strategy is proposed to replace the lengthy, complicated, and cumbersome preprocessing step in order to address the aforesaid issues and propose a high-precision multiclass retina identification system with a deep transfer learning method. Instead of applying contrast-enhancing to the whole fed retina picture database, this suggested solution uses a proposed single convolution neural network to decide the contrast-enhancing for the poor quality of retina photos. Furthermore, a unique picture enhancement filter has been introduced to increase the proposed multiclass model’s output classification accuracy. Finally, a unique augmentation approach has been created to expand the dataset's generality while reducing the impact of a lack of input samples. For each class, the samples were rotated eleven times by 30 degrees. The performance of the proposed model, which is based on fine-tuned and pre-trained ResNet101 and VGG19, was assessed using four databases: FIGSHARE, IDRID, MESSIDOR, and STARE. The FIGSHARE achieved a 100% classification accuracy, 100% sensitivity, 100% specificity, and 100% F1-score for FIGSHARE; while 92.6% classification accuracy, 94.15% precision, 92.62% sensitivity, 97.38% specificity, and 92.40% F1-score were achieved using MESSIDOR. The IDRID achieved a 93.6% classification accuracy, 92.74% precision, 94.05% sensitivity, 98% specificity, and 93.05% F1-score. Lastly, STARE obtained a 99.29% classification accuracy, 99.29% precision, 99.29% sensitivity, 100% specificity, and 99.29% F1-score. In comparison to the state-of-the-art research, the suggested model outperformed them by a significant margin of accuracy, sensitivity, and specificity for the retina classification. Therefore, this study is a promising retinal diseases diagnosis framework, which potentially supports the ophthalmologist in fast and precise diagnosis and treatment to stop vision impairment and loss.