Métodos computacionais para identificação, quantificação e análise de mudanças no tecido da lesão pulmonar através de imagens de tomografia computadorizada

Abstract

Lung cancer is one of the most common types of cancer around the world. Temporal evaluation has become a very useful tool when to whoever needs to analyze a lung lesion. The analysis occurs when a malignant lesion is under treatment or when there are indeterminate lesions, but they are probably benign. The objective from this work is to develop computational methods to detect, quantify and analyze local and global density changes of pulmonary lesions over time. Thus, it were developed four groups of methods to perform this task. The rst identi es local density changes and it has been denominated voxel-based. The second one is composed of the Jensen divergence and the hypothesis test with global and local approaches. Similarly, the third group has only one method, the principal component analysis. The last group has one method, it has been denominated modi ed quality threshold, and identi es the local density changes. In order to reach the objectives, it was proposed a methodology composed of ve steps: The rst step consists in image acquisition of the lesion at various instants. Two image databases were acquired and two models of lesions were created to evaluate the methods. The rst database has 24 lesions under treatment (public database) and the second has 13 benign nodules (private database) in monitoring. The second step refers to rigid registration of the lesion images. The next step is to apply the proposed four groups of methods. As a result, the second group of methods detected more density changes than the fourth group, which in turn, this latter detected more regions than the rst group and this more than the third group, for the public database. For the private database, the fourth group of density change methods detected more regions than the rst group. The third group detected few regions of changes when compared to the rst group and the second group had the lowest number of detected regions. In addition to the density changes found, the proposed classi cation model with texture features had accuracy above 98% in the diagnosis prediction. The results state that there are changes in both databases. However, the detected changes for each group of methods have di erent intensity and location to the databases. This conclusion is based from high accuracy that was obtained from the prediction of the lesion diagnosis from both databases.