Glaucoma Detection and Classification Using Improved U-Net Deep Learning Model

Abstract

Glaucoma is prominent in a variety of nations, with the United States and Europe being two of the most famous. Glaucoma now affects around 78 million people throughout the world (2020). By the year 2040, it is expected that there will be 111.8 million cases of glaucoma worldwide. In countries that are still building enough healthcare infrastructure to cope with glaucoma, the ailment is misdiagnosed nine times out of ten. To aid in the early diagnosis of glaucoma, the creation of a detection system is necessary. In this work, the researchers propose using a technology known as deep learning to identify and predict glaucoma before symptoms appear. The glaucoma dataset is used in this deep learning algorithm that has been proposed for analyzing glaucoma images. To get the required results when using deep learning principles for the job of segmenting the optic cup, pretrained transfer learning models are integrated with the U-Net architecture. For feature extraction, the DenseNet-201 deep convolution neural network (DCNN) is used. The DCNN approach is used to determine whether a person has glaucoma. The fundamental goal of this line of research is to recognize glaucoma in retinal fundus images, which will aid in assessing whether a patient has the condition. Because glaucoma can affect the model in both positive and negative ways, the model's outcome might be either positive or negative. Accuracy, precision, recall, specificity, the F-measure, and the F-score are some of the metrics used in the model evaluation process. An extra comparison study is performed as part of the process of establishing whether the suggested model is accurate. The findings are compared to convolution neural network classification methods based on deep learning. When used for training, the suggested model has an accuracy of 98.82 percent and an accuracy of 96.90 percent when used for testing. All assessments show that the new paradigm that has been proposed is more successful than the one that is currently in use.

Keywords: DenseNet-201 model; classification; deep convolution neural network; image segmentation; improved U-Net.

Figure 2

Photographs of the optic disc showing a normal disc (0) and optics with varying degrees of glaucomatous optic nerve damage (Stages 1–5). These phases are classified based on the shape of the neuroretinal rim (blue arrow) [8].

Figure 1

Glaucoma images: (a) macular epiretinal membrane, (b) normal fundus, (c) mild nonproliferative retinopathy, (d) pathological myopia, (e) hypertensive retinopathy, (f) laser spot, moderate nonproliferative retinopathy, (g) moderate nonproliferative retinopathy, laser spot, (h) mild nonproliferative retinopathy.

Figure 3

The Internal Architecture of the Proposed Model.

Figure 4

The improved U-Net model. Process see Box 1.

Figure 5

An image created by U-Net that is compared to the actual ground-truth Column 1 shows original images, Column 2 shows output images generated by the proposed method, and Column 3 shows ground-truth images.

Figure 6

Feature extraction using pretrained DenseNet-201 model and classification using DCNN.

Figure 7

The internal process of the proposed improved U-Net model.

Figure 8

Comparative analysis of the testing and training accuracy of the proposed model with traditional models.

Figure 9

A comparative analysis of precision of the proposed model with traditional models.

Figure 10

A comparative analysis of recall of the proposed model with traditional models.

Figure 11

Comparative analysis of specificity of the proposed model with traditional models.

Figure 12

F-measure of the proposed model with traditional models.