Avaliação das métricas de autofoco para aplicação em imagens de baciloscopia de tuberculose obtidas utilizando microscopia de campo claro

Abstract

Tuberculosis is considered a world-wide political issue. It has been known that a third of the world population is infected with the Mycobacterium Tuberculosis bacillus and consequently can develop the disease. The finding of the disease is carried by bacilloscopy, which consists of the counting of bacilli in 100 different fields of a slide with sputum sample smear. This type of work demands a lot of training and has a high false negative rate. In order to increase the performance of the bacilloscopy, the development of an automated system to perform the counting of the bacilli is necessary. Our work is based on the study of initial stage of the automated process, emphasizing the autofocus (AF). We investigated nine autofocus functions (Brenner gradient, energy Laplace, wavelet transformation, variance, normalized variance, autocorrelation, standard-deviation-based correlation, variance of log histogram and entropy), widely used in literature, in order to define the best function for images of bacilloscopy of conventional microscopy stained with Kinyoun method. In order to identify the focus functions that had the best performance the focus curve was evaluated by five criteria (accuracy, range, number of false maxima, width and execution time) with two types of analysis (semiquantitative and quantitative). This methodology was performed on images with high and low density of contents, as well as on images with a lot and few amount of bacillus. The total images for the experiment was 1800. Through this study, it was determined that the best function for this type of image is based on Variance. In order to deepen the results, we performed the same study for Kinyoun images without the final step of staining, because it results in a background image more uniform, which can be used for future applications for the automation of counting of bacilli. For these types of images, the function that showed the best performance was based on Entropy.