Estudo de técnicas de segmentação de imagens baseadas em contorno ativo aplicadas em tomografia computadorizada de pulmão

Abstract

In the global rank of leading causes of death are several lung diseases, such as chronic obstructive pulmonary disease; lower respiratory tract infections; and cancer of the trachea, bronchi, and lung. COVID-19 can also be included in this ranking. Thus, early diagnosis is essential to reduce the mortality of these diseases. The segmentation of lung lesions is a way to support the diagnosis that allows the radiologist to better visualize the region of interest and evaluate the progression of the disease. Thus, the present work evaluates two segmentation algorithms based on active contour, performing pre-processing and segmentation techniques for extraction of the lungs and lung lesions from CT images, along with the development of a MATLAB interface for processing selection. The database used for the tests is composed of the selection of lung tomographic slices of exams present in the repository of the Hospital São Lucas of the Pontifical Catholic University of Rio Grande do Sul (HSL-PUCRS) containing tomographic images of lungs with different pathologies. The processing was divided into several pre-processing and segmentation approaches, visually evaluating the detection and segmentation of lung lesions with or without noise smoothing by filters; with or without contrast stretching; and using the active contour segmentation techniques by the MATLAB integrated function, activecontour() (MATLAB ACM), which performs the active contour technique by level-set sparse field, and with the use of the online region ORACM technique, developed by Talu (2013). Overall, the Online Region Active Contour Model technique performed more effectively than the built-in MATLAB function activecontour(), detecting more lesions, achieving fewer extraction errors, and delineating the lesions' shape. Nevertheless, this method could not detect all the lesions present in the image. Furthermore, after comparing different preprocessing, it was observed that the two methods were able to detect the lesions more easily without applying smoothing filters and performing contrast stretching. The area of lesions detected by the two methods was similar in most slices tested, however, there were slices that presented detection errors. Therefore, it is necessary to expand the analysis by evaluating another strategy for extracting the lung area; other analysis and segmentation metrics; as well as other pre-processing techniques. Regarding the interface, it allowed easy interaction with the segmentation and comparison algorithms, also making the code shareable. However, it needs to undergo further optimizations.