Detection of Mycobcaterium tuberculosis in microscopic images of Ziehl-Neelsen-stained sputum smears LACNEM 2015

摘要

Tuberculosis is a disease with a high mortality rate worldwide, but early recognition highly increases the chances of survival. There are several tests to diagnose this disease, however the Pan American Health Organization recommends for its low cost and effectiveness sputum smear microscopy. The count bacilli present in a sputum sample is the core of diagnosis. This procedure requires time and a specialist with a well-trained eye, thus making it prone to error. The techniques used in artificial vision for object detection can detect bacilli, enhancing the performance of bacilloscopies. This paper presents a methodology for detecting bacilli in sputum smear images, implementing adaptive k-mean segmentation clustering, and using artificial neural networks. The database consists in total of 100 images taken from three samples, each with different characteristics. All samples were analyzed by an expert in microbiology, with which the ground truth was obtained. The assessment was done by comparing the segmented image pixel by pixel with the ground truth and calculating the confusion matrix values for specificity, sensitivity, and accuracy. The segmentation succeeded 97.59% precision. The accuracy of the artificial neural networks reached 98% using cross-validation. The results show that the proposed methodology is effective, and it can accurately identify the bacilli in the Ziehl-Neelsen-stained sputum smear images.