Artificial Intelligence-Based Hough Transform with an Adaptive Neuro-Fuzzy Inference System for a Diabetic Retinopathy Classification Model


The influence of the Internet of Health Things (IoT) on the progression of the healthcare industry is enormous and the utilization of artificial intelligence (AI) has transformed IoHT systems at almost every level. In general, diabetic retinopathy (DR) occurs in people who have been suffering from diabetes for a long period and. due to retinal infection, this leads to loss of eyesight. With the application of these models of fundus imaging, the DR-defected retinal structure can be predicted. The fundus images are captured with the application of a fundus camera. The inner surface of the eye is shown by fundus images that are composed with fovea, retina, blood vessels, optic disc (OD), as well as macula. An ordinary retina is composed of blood vessels that are constrained with nutrients and blood supply. The blood vessels are soft and filled with extra blood pressure, and may burst in diabetic patients. By providing additional small blood vessels count, the DR develops owing to the extra pressure exist at the retinal surface. To classify the different stages of DR such as non-proliferative DR(NPDR) and proliferative diabetic retinopathy (PDR) over actual retina, the blood vessels development can be applied as the bio-marker [I].

More developers have been presented with efficient vessel segmentation over the last few decades and retinal images undergo classification based on the disease severity [2, 3]. In the case of DR analysis, an automatic retinopathy classifier has been deployed based on artificial neural networks (ANNs). Genetic algorithm (GA) and fuzzy c-means (FCM) can be applied to attain maximum accuracy value [4]. These models are multilayered thresholding [5] in segmenting blood vessels present in DR images. The retinal structure investigation deals with ridgelet [6], curvelet [7] and wavelet [8] transforms that are utilized with fundus images. Fuzzy logic (FL) is employed to provide higher sensitivity value. With the application of a multiscale line detecting device, the retinal vascular can be analyzed [9]. The combination of a Gaussian mixer and nearest neighborhood approaches is employed in [Ю], who deployed a scheme named DR analysis by applying machine learning (DREAM) and processed the classification task by using support vector machine (SVM). An L2 Lebesgue integral model was applied to compute the infinite perimeter regularization [II]. To preprocess the image, the global thresholding approach has been employed [12]. The blood vessel prediction by morphological component analysis (MCA) reached maximum accuracy value [13].

The technique of deep neural networks (DNNs) [14] has been used, which undergoes training across a greater number of sample STARE, CHEST, and DRIVE data sets. However, it has a manageable number of samples. For system-based screening, a telemedicine system has been deployed by utilizing red lesions of retinopathic images. In the case of DR referral [15], a direct technology is presented by classification training. The reduction of lesion-to-lesion prediction requires numerous pools to perform the classifier training. Many traditional methods such as morphological gradients, wavelets, NN. and alternate computation have to be executed to provide efficient and accurate DR prediction at earlier phases. However, these techniques are more tedious to be computed.

Automatic DR detection techniques are composed with various benefits such as DR can be forecasted at primary stages in an effective manner. Some of the methodologies such as deep learning (DL) have led to the evolution of computer vision. The image classification task is performed by convolution neural networks (CNNs). Research of this application contributes in feature segmentation as well as blood vessels [16]. The actual image classification can be processed by the application of deep CNNs (DCNNs) by classifying DR fundus images. To resolve the issues involved in segmentation of blood vessel, a CNN method is employed in [17] to retrieve image features. Therefore, it is comprised with few limitations of existing methodologies. Then, data sets that have minimum quality and lower fundus images with individual collection platform provide few difficulties for comparing the model’s function. In order to enhance the working function of a method, AlexNet is devised.

Additionally, the qualified models of CNNs are GoogleNet and VGGNet. Recently, the projected Residual Network (ResNet) is assumed to be a more vital system that increments the CNNs function at the time of processing image classification. To increase the learning duration and compare with traditional approaches namely, VGGNet, AlexNet, and GoogleNet, it is employed with transfer learning that provides exact as well as automated prediction with visual infections to be lowered at the minimum degree [5, 18]. Among other models, the presented approaches are composed with consecutive improvements on convergence time for a huge-sized data set and shows qualified function rather than classification.

In this chapter, it is deployed with the automatic segmentation-centric classification method for DR. At this point, contrast-limited adaptive histogram equalization (CLAHE) has been employed for preprocessing and the watershed technique is used in image segmentation. Then, the Hough transform (HT)-based feature extraction process is carried out by the adaptive neuro-fuzzy inference system (ANFIS) method utilized in image classification. In experimental analysis, the data set has been derived from the Kaggle website, which is assumed to be an open-source environment that tries to develop the DR detecting approach.

Overall process of the presented model

FIGURE 11.1 Overall process of the presented model.

< Prev   CONTENTS   Source   Next >