Venkataraman, Karthikeyan2020-09-142020-09-142019-10Venkataraman, Karthikeyan, (2019) Application of computer vision and machine learning for classification of Mycobacterium Tuberculosis (TB) cultures to assist disease diagnostics, University of the Witwatersrand, https://hdl.handle.net/10539/29627https://hdl.handle.net/10539/29627A Research Report submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in partial fulfilment of the requirements for the degree of Master of Computer Science by Coursework and Research ReportDiagnosis of Tuberculosis (TB) by the analysis of sputum samples is a manual activity that requires the sputum culture samples to be analysed after a gestation period between 6 to 12 weeks following the collection of the sample from the patient. The current method of TB diagnosis at the National Heath Laboratory Service (NHLS) is a manual process that involves photographing the culture samples by the laboratory technician using a camera phone, among other steps. It is proposed to use these photographs to automate the process of diagnosis by using Computer Vision and Machine Learning technologies to detect the presence of Tuberculosis bacteria. A photograph of a culture sample or plate, consists of 48 unique circular regions or wells, each of which is classified by the machine learning program as Positive, Negative, Contaminated or Condensation (indicating presence of water bubbles). This classification of each well is done using a combination of Image Processing, Computer Vision and Machine Learning (ML) techniques. Two ML programs, in particular, were used for this task - Gaussian Mixture Models (GMM) and Random Forest (RF). At the time of writing of this report, a classification accuracy of 77.6% using 25 training images was obtained. The results from each ML technique are compared and approaches to further improve the classsification accuracy are discussed. A data flow and machine learning pipleline was developed in Python 3 and the code has been opensourced on GitHub at https://github.com/karthik111/TB-culture-classification-ML allowing future work to be undertaken. The Image Processing and Computer Vision techniques implemented here are agnostic of the downstream ML technique allowing other ML approaches to be explored. This work establishes the feasibility of an automated approach for classification of sputum samples to aid TB diagnosis and describes approaches that if taken, can further improve the classification accuracy.Online resource (ix, 50 leaves)enMycobacterium tuberculosisTuberculosis--DiagnosisThesis