Abstract:

The advancements in medical imaging and information technology have paved the way for the use of computers in diagnosis. It has also become essential to reduce the mortality rate by early detection of diseases. Computer-Aided Diagnostic (CAD) systems are being developed to diagnose diseases from different medical imaging modalities. Interpretation of these images presents challenges to the expert as the affected regions’ patterns may look similar for one or more diseases. CAD systems have been developed to diagnose the disorders present in the lungs, brain, liver, and spinal cord. Physicians can use CAD systems as a second opinion in decision making and treatment planning.

Complementary and Alternative Medicine (CAM) techniques for the diagnosis are also prevalent. In the modern era of computers, medical practitioners believe that the combination of CAM techniques with CAD can accomplish medical science. Iridology is a CAM technique which is generally based on the concept of neural pathways between the body and the iris. The iridologist assessed the health condition, which interprets patterns, shapes, rings, colors and pigmentation markings, fibers, structures, and changes in the iris. In this thesis, an attempt is made to correlate the specific areas of the iris with diabetes. The related disease condition of the individual as diabetes is a chronic disease that occurs either when the pancreas does not produce enough insulin or when the body cannot effectively use the produced insulin. Insulin is a hormone that regulates blood sugar. Hyperglycaemia or raised blood sugar is a common effect of uncontrolled diabetes and, with time, leads to severe damage to many of the body’s systems, especially the nerves and blood vessels.

This research aims to develop an efficient CAD system for automatic screening of diabetes and related diseases with high accuracy using machine learning such that an alternative diabetes diagnosis method can be proposed. Image-based CAD systems generally consist of seven phases-database creation, image pre-processing, segmentation, finding region of interest, feature extraction, optimizing the features, and classification for decision making.

For this research work, we have developed our data-set of infrared eye images along with meta-data. Meta-data includes the diabetes clinical blood test reports, details of all the physiological parameters, and patients’ medical history. A supervised learning model for the diagnosis of diabetes can be trained. The success of a computer-based system depends both on the features and classification method. An efficient set of textural features decides the accurate diagnosis, and an appropriate classification method provides the potential to produce a correct classification. In the first research work, the diagnosis of diabetes is made based on the first-order statistical features and discrete wavelet features from the specific areas of the iris. The investigation was performed over a close group of 338 subjects (180 diabetic and 158 non-diabetic).

According to the iridology chart, the region of interest from the iris image was cropped as the zone corresponds to the position of the pancreas organ. Statistical, textural, and discrete wavelength transformation features were extracted from the region of interest. The results show the best classification accuracy of 89.63% calculated from random forest classifier. Maximum specificity and sensitivity are absorbed as 0.9687 and 0.988, respectively. Results have shown the proposed model’s effectiveness and diagnostic significance for a non-invasive and automatic diabetes diagnosis.

The second work presents a detailed comparative analysis of classification techniques to diagnose type 2 diabetes and its duration using the combination of features of a specific area of iris and physiological parameters. A set of 334 subjects is investigated. Subjects were divided into diabetic and non-diabetic groups. Moreover, the diabetic group was classified into three different subgroups according to the diabetic state’s duration. Statistical features, Gray Level Co-occurrence Matrix (GLCM), and Gray Level Run Length Matrix (GLRLM) based features were extracted from the specific areas of iris. Nine classifiers of different application areas were selected. Subsequently, six parameters (accuracy, precision, sensitivity, specificity, F-score, and area under the curve) of each classifier have been computed and analyzed.

The analysis provides promising results with more than 95% accuracy. Similarly, the same features have been further explored for another study to diagnose Diabetes Kidney Disease (DKD). This experiment achieved an accuracy of 94.5% with the combination of first-order statistical features, GLCM, and GLRLM based features. In the previous study, the sclera-based person identification technique presented is a relatively new biometric technique and needs to explore more.

The human eye’s sclera contains the unique blood vessel patterns that make it a potential tool for personal identification. A fast and robust sclera segmentation algorithm and an impressive feature set for recognition are presented. The most versatile data-set in image quality (UBIRIS V1) was selected for segmentation and recognition purposes. For sclera segmentation, an unsupervised algorithm is presented based on pixel mapping and cauterized grayscale. After that, from the segmented sclera, global features and Gabor wavelet transform based are extracted. Finally, for the recognition, a multi-class support vector machine classifier is applied. The recognition results are presented in terms of overall system accuracy and sensitivity. It is notionally proved that the proposed technique is reliable and accurate for executing sclera-based recognition.

Doctoral Thesis, Sign up to request full thesis:

Thapar Institute of Engineering & Technology