Quantitative Approach to Automated Diagnosis of Malaria from Giemsa-Thin Blood Stain using Support Vector Machine

Abstract

Automating the process of malaria diagnosis is very crucial; malaria is a deadly disease with an annual infection rate between 300–500 million and a death rate of 1 million yearly. The diagnosis approach is manual and is subject to human error. In this current work, we automate the process of diagnosis and provide results in quantitative form with a diagnostic tool deployed on a web server to eradicate limiting the access to the service to a physical location. The input to the developed diagnostic tool is a Giemsa-stain blood image which undergoes image processing using Otsu segmentation to identify regions of the red blood cells, and a trained SVM classifier iterate through the red blood cells to determine the infected ones. The trained SVM achieved accuracy and precision of 88% and 87% against the validation dataset. The count of infected red blood cells against total red blood cells in the image is used to compute the quantitative result which is the level of severity and number of infected cells per uL of blood, based on the World Health Organization (WHO) standard. A couple of Giemsa-stain blood images were uploaded for diagnosis, our web-based diagnostic tool achieved 90.55%, 85.7% and 100% for average count (both total red blood cells and total infected red blood cells in the processed Giemsa-stain blood images) accuracy, severity classification accuracy and negative test accuracy respectively. The system's average time to complete a diagnosis is 2.2824 seconds, this is a very short time which will create a near-real-time experience for the users of the service.