Uma Nova abordagem para a segmentação pulmonar com reinclusão de nódulos

Abstract

In several applications of medical image analysis, the image segmentation process is a fundamen- tal task, especially when it comes to computer systems aimed at assistance in terminal diseases diagnosis, such as lung cancer. Thus, one of the major challenges lies in the segmentation accuracy regarding the preservation of internal structures, especially the just-pleural pulmonary nodules, which are not usually included in the final lung segmentation. This missing information may be critical in inspecting structures and identifying diseases, monitoring evolution, or even simulations for surgical planning. That way, different approaches are used to promote reinclusion of the nodules, therefore, in the final solution of the problem. In this sense, this dissertation proposes a new approach for pulmonary segmentation in computed tomography images of the chest with nodules reinclusion using classical mathematical morphology techniques to reduce nodule loss. This approach consists of a preprocessing sequence, followed by morphological pulmonary segmentation by 3D connected component analysis, and non-morphological hole filling, and finally the post-processing which main purpose is the reinclusion of the lost nodes, and has its process divided into two stages, namely the closure of the lateral pulmonary pleura through the its convex hull area, and the morphological closure of the mediastinal region. To validate this approach, tests are performed in an extensive database (LIDC-IDRI) as well as a gold standard generated with the assistance of medical specialists and compared to a classical segmentation technique, 3D Region Growing, and a classification-based method found in the literature. The main contributions of this dissertation are in the fill holes technique, which replaces the morphological method for greater computational efficiency, as well as the developed developed for reinclusion of justa-pleural nodules. The results prove that there is a substantial gain in processing time in the holes filling step, which is six times faster than the morphological counterpart. In addition, the proposed segmentation achieves a minimum loss of 1.9 % of nodules from a total amount of 2663 nodes, and reaching 91 % and 87 % for Similarity Coefficient and Fitness Adjust rates, respectively, in comparison of the final segmentation with the gold standard. In this context, it is concluded that the proposed segmentation, according to the metrics used, presents results superior to the methods compared to the loss of nodules, as well as the reduced processing time, being compatible with the methods found in the literature regarding measures of segmentation quality.