Name: LUCAS DE ASSIS SOARES
Type: MSc dissertation
Publication date: 17/04/2015
Advisor:
Name |
Role |
|---|---|
| EVANDRO OTTONI TEATINI SALLES | Co-advisor * |
Examining board:
| Name |
Role |
|---|---|
| EVANDRO OTTONI TEATINI SALLES | Advisor * |
| KLAUS FABIAN COCO | Internal Examiner * |
| PATRICK MARQUES CIARELLI | Internal Examiner * |
Summary: Even though tuberculosis (TB) is a treatable disease, it is still a major global health
problem being second only to AIDS as the greatest killer worldwide due to a single
infectious agent (WHO, 2015). In order to treat it, the disease must be properly diagnosed.
The diagnosis is usually done using by staining a slide with patient sputum using the
Ziehl-Neelsen stain and then a human specialist analyzes it using an optical microscope
looking for tuberculosis bacilli. Since this process is time consuming and labor intensive,
an automatic bacilli recognition system allows the diagnosis process to be more agile and
less tiresome. In this work, an automatic tuberculosis bacilli segmentation system using
conventional microscopy images is proposed. The system is basically divided in two stages:
a stage of segmentation and a stage for classification of the segmented structures. First,
images are projected based on a linear discriminant analysis considering Fisher criterion
in order to increase the separation between bacilli pixels and background pixels. Then,
two approaches are evaluated: a segmentation process based on global thresholding and
another based on Otsu segmentation method. Structures that have a big or a small area
are then filtered and morphological operators are applied over the binary image. Finally,
the segmented structures are classified using artificial neural networks and support vector
machines. The results show that it is possible to implement an automatic tuberculosis
bacilli segmentation system that provides a good distinction of bacilli in the images with
a low computational cost. For the segmentation stage, up to 98.69% of bacilli are correctly
segmented and up to 85.51% of bacilli remain after the area filter. For the classification of
the structures, mean values up to 94.25%, 95.33%, 95.73% and 92.50% were obtained for
sensitivity, specificity, accuracy and precision, respectively.
