A Machine Learning System for Detection and Grading of Diabetic Maculopathy in Retinal Oct Images

Abstract

Maculopathy is a collective group of diseases that damages the central region of a retina known as macula. A common form of damage is from Diabetic Macular Edema (DME) in which fluid builds up on the macula. The risk of this problem can be completely prevented by means of two fundamental public health interventions such as early diagnosis and treatment. However, the effective screening and diagnosing the DME from the retinal Optical Coherence Tomography (OCT) images is a major challenge. Optical Coherence Tomography (OCT) constitutes an imaging technique that is increasing its popularity in the ophthalmology field, since it offers a more complete set of information about the main retinal structures. Hence, it offers detailed information about the eye fundus morphology, allowing the identification of many intra retinal pathological signs. In this scenario, the experts and other professional diagnosis teams depend on analyzing the retinal OCT images using the Computer-Aided Diagnosis (CAD) system. The general processing stages of CAD system involves preprocessing, segmentation and classification pertaining to retinal abnormalities. Various machine learning methods have been developed. Further the recent contribution of deep learning models and its successful performance over the conventional techniques under medical applications have been motivated the researchers for adopting the deep learning models to diagnose the DME. The classification of DME is critical in clinical decision-making because it defines the stage and thus the severity of the disease. Hence, this thesis propagates towards detection and the grading of the Diabetic Maculopathy(DM) patients based on their retinal pathological conditions, anatomical histology and aims to develop diagnostic strategy and assistive tools for automated classification. newlineThe main aim of the research is to effectively segment the layers of retinal OCT images and carry out the classification of the DME. In order to achieve the goal a novel technique is proposed to carry out segmentation and Classification. The thesis explains the proposed Diabetic Macular Edema (DME) segmentation and classification techniques in four contributions. For the first contribution, goal of the research is to design and develop a Gradient-based Adaptive thresholding with Active contour model for the detection of DME. Initially, the input image is given to the layer segmentation module where the layer segmentation is carried out based on Gradient-based Adaptive thresholding integrated Active contour model. In addition, the performance of the proposed method is implemented with the OCT Image Retinal Dataset and analyzed using three performance metrics, namely segmentation accuracy, dice coefficient, and Jaccard coefficient. In the second contribution, Ant Lion Spider Monkey Optimization-based Generative Adversarial newlineNetwork (ALSMO-based GAN) is invented to segment and classify the retinal layer for identifying the DME. In the third contribution, Gradient-based Adaptive Thresholding integrated Active Contour Ant Lion Spider Monkey Optimization driven General Adversarial network (G-AT_AC ALSMO-GAN) is introduced for classifying the DME from OCT images. newlineIn the fourth contribution, Shape Index Histogram Honey Badger Aquila Optimization-based deep convolutional neural network (SIH+HBAO-based deep CNN) method was invented to detect the DME. Here, feature extraction is performed whereas layer specific features, texture features include Local Gradient Pattern (LGP) and proposed SIH with multi-kernel are extracted. After that, DME detection and classification is conducted utilizing Deep CNN, which is tuned employing proposed HBAO algorithm. Howsoever, proposed HBAO algorithm is introduced by an incorporation of Honey Badger Algorithm (HBA) and Aquila Optimizer (AO). Moreover, the implementation of proposed DME segmentation and classification techniques are done by the python tool using OCT Image Retinal Dataset. In addition, the effectiveness of proposed model is analyzed with the performance metrics, such as testing accuracy, sensitivity and specificity. For the DME segmentation, the G-AT+active contour attained the segmentation accuracy of 91.55%, Jaccard coefficient of 0.9155 and Dice coefficient of 0.5081 with salt and pepper noise and G-AT+active contour attained the segmentation accuracy of 92.35%, Jaccard coefficient of 0.9235 and Dice coefficient of 0.5292 for the Poisson noise. Additionally, in the first level classification, the testing accuracy, sensitivity and specificity of SIH+HBAO-based Deep CNN is 0.9117, 0.9170 and 0.9181, whereas the second level classification attained the evaluation metrics of 0.9067, 0.9120 and 0.913. newlineKeywords: Diabetic Maculopathy, Diabetic Macular edema, Optical Coherence Tomography, Active Contour, Gradient-based Adaptive Thresholding integrated Active Contour, Honey Badger Algorithm, Aquila Optimizer. newline