The development of artificial intelligence system to detect Mycobacterium tuberculosis (M. tuberculosis) from sputum with Acid-Fast Bacillus (AFB) method

Abstract

Pulmonary tuberculosis is an infectious disease that still has outbreaks in Thailand (Department of Disease Control, 2021), caused by a bacterium called Mycobacterium tuberculosis, often known as MTB. In Thailand, the AFB sputum smear test is commonly used to diagnose tuberculosis. It is reliable and affordable, however the physician inspection process is time-consuming. So, we created an AI to automatically inspect the sputum smear to support the physician's inspection procedure in which we started by collecting sputum images from Kaggle and ZNSM-iDB, and processed them through data cleaning and augmentation process. Then, we developed the program for supporting all of the AI, coupled with the Yolov5 module to train, evaluate, and finally compare the performances of the different CNN models used in Yolov5's image processing process. As a result, the system is able to learn from the image dataset, which in preliminary examination, which classifies between positive and negative sputum sample, the most effective model, Yolov5s, was able to detect the MTBs with the sensitivity, specificity, F1-score, and mean detection time per image of 0.9802, 0.9647, 0.9727 and 11.1 milliseconds, respectively, and in the object detection task in which locates every TB in the sample, the most efficient model, Yolov5n6, was able to detect objects and classify the types of MTB with the precision, recall, mAP of 0.673, 0.761, 0.727 respectively. Once the model is exported, the CNN image processing model is trained and ready to be used in further experiments or further development into various innovations. To conclude, the model is able to inspect the MTBs automatically with a high accuracy and fast detection time. However, model image predictions in some images are still wrong, possibly because of the different staining characteristics in each picture, causing the system to misinterpret.