Evaluating Semantic Segmentation of Tuberculosis Bacilli in Bright Field Microscopy Using Different Color Spaces Components and Mosaic Images

Tuberculosis is one of the infectious diseases that causes the most victims in the world. The early diagnosis of the disease is fundamental for the treatment to be carried out quickly, saving lives and decreasing the number of infected people. The tuberculosis diagnosis is a time-consuming process that involves the analysis of up to 100 fields of conventional microscopy, which causes fatigue for the technician involved in the process. In order to assist the technician in counting bacilli in tuberculosis bright field microscopy examinations, automatic methods using deep neural networks are proposed. In this work, we evaluate the use of a semantic segmentation network for bacilli detection, using as input several color components from the RGB, HSV, YCbCr, and Lab spaces. The model that presented the best performance uses only the components of RGB color space as input to the network, with values of accuracy, precision, sensitivity, specificity, and f1-score above 99%.