Quantitative Analysis of Pathological Strips using Machine Learning

Abstract

Urinalysis is a significant technique used for inspecting the urinary system. Urine has numerous chemical materials secreted; these materials can be used to diagnose diseases and conditions such as urinary tract infection, diabetes, kidney diseases, and pregnancy, at the earliest. These diseases affect people of every age, gender, and profession. If not diagnosed timely it can cause death in a matter of months if not days. Mostly, urinalysis is used for diagnosing it, but studies show that it is a costly procedure and is prone to human error. Moreover, numerous studies and research works have been carried out to present techniques for diagnosing these diseases, and they have achieved success to a great extent. However, there is still room for improvement, and until 100% accurate results are achieved, the struggle for diagnosing and treating diseases would continue. This thesis has conducted a systematic literature review to thoroughly analyze the existing literature and look for gaps in it. 33 research studies that have been published during 2007-2019 related to the specified domain have been identified and analyzed. This leads to the identification of 10 methods, 8 technologies, 12 challenges, and 10 diseases. The main objective of this thesis is to present a technique that can be used to develop an automated system for urinalysis. The proposed technique has been able to outclass the previously presented techniques in terms of accuracy. It has been able to achieve the highest accuracy because of the different experimentations carried out to analyze their results. Besides this, the results of the proposed methodology have also been evaluated and compared with one of the previously used methods i.e. Euclidean Distance. The dataset used in this manuscript contains the urine test strip RGB images, which is developed using MATLAB. Different noises i.e. Gaussian and salt & pepper have been applied on the dataset (images) to increase the number of instances. Initially, the dataset was comprised of three regent color strip images, which are used as inputs for Euclidean distance. After this, segmentation has been applied to the images and then the segmented images have been used for experimentation purposes. The accuracy vs noise variance graphs for the models have been presented in Figure 5-10. The experimentation results reveal that the regression model with the CNN classifier has a higher accuracy as compared to Euclidean Distance.