Retinal Disease Diagnosis through Computer Aided Fundus Image Analysis

Abstract

Diabetic retinopathy, an asymptomatic complication of diabetes, is one of the leading causes of blindness in the world. The early detection and diagnosis can reduce the occurrence of severe vision loss due to diabetic retinopathy. Therefore, the present research work was conducted to diagnose symptomless clinical stages of diabetic retinopathy, i.e., non-proliferative diabetic retinopathy and progressive diabetic retinopathy. The diagnostic confirmation of diabetic retinopathy depends on the reliable detection and classification of bright lesions namely: exudates and cotton wool spots, and dark lesions namely: microaneurysms and hemorrhages present in retinal fundus images. However, variability within the retinal images makes difficult to distinguish dark and bright lesions in the presence of anatomical structures (landmarks) like blood vessels and optic disk. Therefore, it is necessary to eliminate any spurious, unwanted and false regions due to anatomical structures before the segmentation of retinal lesions. In addition, to design an efficient computer-aided diagnostic method a benchmark composite database, having varying attributes such as position, dimensions, shapes and color is required. Keeping all these facts in mind, a composite database has been designed in this work to provide an efficient and generalized computer-aided solution for the diagnosis of diabetic retinopathy In this research work, a composite database formed, includes 5048 retinal fundus images from two diverse sources: one is clinical source including 2942 images and other is six online available benchmark sources including overall 2106 images. The clinical database was developed by acquiring database of 2942 retinal images with varying color, brightness and quality from 482 different patients: 280 men (mean age: 51 years) and 202 women (mean age: 44 years) with an age range of 25 to 83 years over the period of January 2014 to July 2017. Image database comprises of 99 healthy and 2843 pathological retinal images with varying grades of diabetic retinopathy. Pathological retinal images comprise of 1351 images with hard exudates, 446 images with cotton wool spots, 1165 images with micro aneurysms and 1588 images with hemorrhages. The open-source benchmark databases of retinal images used in the present work are DRIVE, STARE, e-OPHTHA, MESSIDOR, DIARETDB and HEI-MED. These databases have different goals, characteristics and levels of completeness. Therefore, a framework for the development of benchmark composite retinal fundus image database is also designed. Artifacts and blur are the prime factors which degrade the contrast resolution and obstruct the meaningful information present in retinal fundus image. Their presence severely hampers the interpretation and analysis of retinal fundus images. Therefore, retinal fundus image enhancement is designed in the present work with the intention to improve the visualization of details available in retinal fundus images in order to provide accurate detection of anatomical structures and retinal lesions. This is accomplished by retaining high frequency components and enhancing the local contrast in retinal fundus images. Two-directional high pass filter is applied to retain high frequencies and then adaptive scaling factor is used for contrast enhancement. Scaling factor is adaptive in the sense that it chooses two different scaling values depending on local variations in image intensities. A threshold value is optimized to make a choice in between two scaling values by analyzing the extent of range of intensity variations in each retinal image database. The enhancement of the images was approved and confirmed by the visual interpretations of expert ophthalmologists. The precise segmentation of blood vessels, in order to avoid spurious responses, is the first step in extracting diagnostic information for the early detection of diabetic retinopathy. Therefore, in the present work a generalized method to detect and segment blood vasculature using retinal fundus images has been designed with four phases namely, (i) initial segmentation of vasculature map to find vessel and non-vessel structures, (iii) extraction of relevant set of geometrical based features from the vasculature map and intensity based features from original retinal fundus image that differentiate vessel and non-vessel structures efficiently, (iv) supervised classification of vessel and non-vessel structures using the extracted features, and (v) joining of candidate vessel structures to create connectivity. Results of subjective evaluation are supported by the objective medically significant statistical measures. The performance of the proposed method has been validated on composite database of 5048 images. The experimental results indicate the high performance of the proposed method with the overall average sensitivity of 84.82% revealing its ability to perform significantly in distinguishing true vessel structures from non-vessel structures. The segmentation results by the proposed method also show a high correlation with the ground truth with an accuracy of 97.58%. Furthermore, the method proves its capability on varying grades of retinal fundus images with the sensitivity of 85.13%, 85.73% and 82.41% on mild, moderate and severe diabetic retinopathy respectively. Finally, it can be emphasized that the promising results indicate the generalization ability of the blood vessels segmentation method to aid ophthalmologists in precise detection of retinal lesions for timely treatment of retinal abnormalities. After the retinal blood vessels are segmented, they are used to segment another anatomical structure i.e. optic disk. In this work, optic disk is segmented in two-steps, viz., (i) localization using blood vessels convergence-based approach and (ii) boundary estimation using canny edge detector and iterative circular Hough transform. After the blood vessels and optic disk are segmented, they are eliminated from the original retinal fundus image. In order to eliminate blood vessels from original retinal fundus image, an adaptive blood vessels normalization method is designed in this work. Subsequently, the optic disk is also eliminated from normalized blood vessels image using morphological filling operation. The exudates, abnormal leaked fatty deposits on retina, are one of the most prevalent clinical signs of progressive diabetic retinopathy. Therefore, computer-aided detection of progressive diabetic retinopathy requires the quantitative assessment of exudates in retinal fundus images. Additionally, during mass screening of diabetic retinopathy, it is vital to differentiate retinal images purely based on the presence or absence of exudates. Therefore, proper detection and then segmentation of exudates is important for decision making during treatment. In this work, a generalized exudates segmentation method is designed by introducing an adaptive image quantization that reduces the number of distinct colours in an image by grouping the image pixels with similar attributes. The proposed method is adaptive in the sense that it adapts itself to the variations in intensity ranges of retinal fundus images in the composite database. In the next step, dynamic decision thresholding method is designed to discriminate pixels corresponding to true exudates from other regions of retinal image. Three parameters were optimized in the designing of decision threshold method by conducting a series of experiments with the help of expert ophthalmologist in order to achieve best combination of sensitivity/specificity. The main contribution of the proposed method is that it reliably segments exudates irrespective of associated heterogeneity, bright and faint edges. The detection capability of the proposed method is assessed using two evaluation criteria: (i) image-based evaluation and (ii) lesion-based evaluation. Image based evaluation criterion evaluates the performance of the method in discriminating images with or without exudates. Whereas, lesion-based evaluation compares the segmentation results pixel by pixel with the reference ground truths. Experimental results for lesion-based evaluation show that the proposed method outperforms other existing methods with a mean sensitivity/specificity/accuracy of 87.95/96.45/92.64 on a composite database of 5048 retinal fundus images. The segmentation results for image-based evaluation with a mean sensitivity/specificity/accuracy of 93.25/97.64/97.48 respectively prove the clinical effectiveness of the method during screening. Experienced ophthalmologists diagnose and then grade non-proliferative diabetic retinopathy by visualizing various shape, intensity and texture-based features of retinal lesions present in retinal fundus images. Therefore, in this work, an extensive feature set is formulated based on different descriptors such as shape, size, intensity, texture of both dark and bright lesion to design an efficient computer-aided severity level detection of non-proliferative diabetic retinopathy. A set of 54 features were extracted using three methods namely: (i) geometrical features based on shape, size and contour of the retinal lesions, (ii) textural features namely: first order statistical (FOS) texture-based features, grey level run length matrix (GLRLM) texture-based features, and grey level co-occurrence matrix (GLCM) texture-based features, and (iii) RGB and HSI model-based colour features. Furthermore, in order to achieve high classification accuracy, the proposed method is designed using a two-step neural network classifier. The main contribution of the proposed method is that it reliably judges the severity level by segmenting dark and bright lesions present in retinal fundus images regardless of types of lesions, blurred and well-contrasted lesions. The experimental results on a composite database of 5048 retinal images for lesion-based evaluation reveals high performance of the proposed method in the segmentation of dark and bright lesions with sensitivity/specificity/accuracy of 94.80/99.80/98.43 and 95.80/98.80/95.43 respectively on clinically acquired database retinal images. The image-based evaluation outcomes show the high precision in precise recognition of clinically acquired pathological images from healthy retinal images with average sensitivity/specificity/accuracy of 98.8/100/100 and 100/100/100 for dark and bright lesions respectively. The overall grading accuracy of 96.33% reveals the capability of the proposed framework in the efficient grading of severity level of non-proliferative diabetic retinopathy based on the detection of dark and bright lesions which are subtle and tough to distinguish. Furthermore, it can be emphasized that the proposed severity level detection method will aid ophthalmologists for appropriate treatment and effective preparation in the diagnosis of various stages of non-proliferative diabetic retinopathy. Finally, it can be concluded that the present research work comprises of (i) retinal image enhancement, (ii) segmentation and elimination of anatomical structures, (iii) detection of progressive diabetic retinopathy and, (iv) detection and classification of non-proliferative diabetic retinopathy. The substantial combined performance of different experiments on clinical and open-source benchmark databases proves the generalization ability and the strong candidature of the proposed computer-aided methods in real-time diagnosis of progressive and non-proliferative diabetic retinopathy.