Retinal Disorders Detection and Analysis From Fundus and Optical Coherence Tomography Images Using Deep Learning Models

Abstract

A brilliant vision gives us a luscious life. The retina of human eye plays a signif- icant role in vision. Damage to any part of retina leads to visual impairments or total blindness. Diagnosis of retinal disorders is a daunting task for ophthalmologists as the devices are not equipped with automatic retinal analysis. The advent of deep learn- ing has transformed the necessity of smart medical applications to a reality in recent years. However, the existing indigenous automatic retinal disorder detection system is designed to grade a single retinal disorder such as Diabetic Retinopathy. Motivated by this, the proposed research aims to identify retinal disorders from multiple retina imaging modalities, namely fundus and Optical Coherence Tomogra- phy images. As fundus and Optical Coherence Tomography images differ in terms of image acquisition procedure, different artifacts affect the image quality. Adapting to the significant difference in image quality in these two modalities, two novel prepro- cessing approaches are proposed in Chapter 3 of the thesis. Histogram is used along with statistical analysis to assess the quality of the acquired retinal images. The pro- posed retinex based non-local total generalized variation restoration method enhances the fundus images increasing the visibility of the macula region. Realizing the fact that the speckle, inherent in Optical Coherence Tomography images are multiplicative in nature, a statistical analysis is incorporated to identify appropriate noise distribution. The non-local deep image prior, discussed in the thesis, despeckles Optical Coherence Tomography images eliminating the requirement of a large number of ground truth im- ages for denoising. The proposed mathematical model using Bayesian MAP estimator and variational models are assessed through implementation. The qualitative and quan- titative analysis presented in Chapter 3 of the thesis confirms that the proposed method outperforms other existing methods. The proposed model restores the image quality while retaining the edge and texture details in the image. Particularly, metrics such as Equivalent number of looks and entropy plots demonstrate that the proposed image restoration model works better than the other existing techniques. Variants of Generative Adversarial Networks are proposed to classify the input reti- nal images into normal or abnormal categories. The abnormal categories include Age- i related Macular degeneration, Glaucoma, and Diabetic Macular Edema. Multiple pub- licly available repositories are preprocessed as described in Chapter 3 of the thesis. The preprocessed images are utilized to train the model and the results are presented in Chapter 4 of the thesis. Simultaneous segmentation and classification tasks are per- formed where the segmentation includes blood vessel extraction, optic disc region and fovea region extraction from fundus images. The performance of Generative Adversar- ial networks for various tasks such as segmentation and classification of retinal images is analysed in Chapter 4 of the thesis. The experimental analysis shows classification accuracy of upto 90% can be achieved proving the stability of a GAN amongst hetero- geneous datasets. Other classification metrics such as F1-score and sensitivity are used to compare the proposed GAN model with other deep learning models. In short, the thesis provides deeper insight into predominant retinal disorders, imag- ing modalities, existing state-of-the-art works on the retinal image analysis through Chapters 1 and 2. The significant contributions of the thesis are discussed in Chapters 3 and 4. Finally the conclusion and scope of future work is presented in Chapter 5. This research work acts as a cornerstone in developing an end-to-end standalone application.