SFPT neural: nova técnica de segmentação de fissuras pulmonares baseada em texturas em imagens de tomografia computadorizadas do tórax

Abstract

Among all cancers, lung cancer (LC) is one of the most common tumors, an increase of 2% per year on its worldwide incidence. In Brazil, for the year of 2014, 27,330 new cases of LC are estimated, these being 16,400 in men and 10,930 in women. In this context, it is of fundamental importance for public health the identication on early stages of lung diseases. The diagnosis assistance shows to be important both from a clinical standpoint as in research. Among the factors contributing to this scene, one important is the increasing accuracy of diagnosis of a medical expert as you increase the number of information about the patient's condition. Thus, certain disorders might be detected early, including saving lives in some cases. The initial treatment for this disease consists of lobectomy. In this context, it is customary to perform the segmentation of lung lobes in CT images to extract data and assist in planning for lobectomy. The segmentation of the lobes from CT images is usually obtained by detection of pulmonary fissures. Thus, in order to obtain a more effective segmentation of pulmonary fissures, and perform a completely independent process from the other structures present in the CT scan, the present work has the objective to perform the fissure segmentation using LBP texture measures and Neural Networks (NN). To implement the algorithm we used one MLP with 60 inputs, 120 hidden neurons and 2 output neurons. The input parameters for the network was the LBP histogram of the voxel being analyzed. For network training, it was necessary to create a system to label the features as fissures and non-fissures manually, where the user selects the fissure pixels class. To perform the validation of the algorithm was necessary to create a "gold standard"in which it was extracted a total of 100 images from 5 exams from the dataset LOLA11, where these images were the fissures were highlighted by two experts. From the gold standard, the proposed algorithm was processed and the results were obtained. For all tested images, the classifier obtained a better performance when the size of 15x15 pixels of the window was used to generate the histogram of the LBP. To get to this definition were tested sizes of 11x11, 15x15, 17x17 and 21x21 and the results were based on metrics comaprados ACC (%), TPR (%), SPC (%) distance mean and standard deviation of the distance. The first approach to analyze the results is through the voxels defined as fissure at the end of the proposed methodology. For the proposed methodology, using automatic detection and MLP LBP before thinning, the rates were obtained ACC= 96.7 %, TPR = 69.6 % and SPC = 96.8 % and ACC = 99 2 % TPR = 3 % and SPC = 99.81 % for the proposed method with the thinning in the end, considering the incidence of false positives and false negatives. Another approach used in the literature for evaluating methods of fissure segmentation is based on the average distance between the fissure delineated by the expert and the resulting fissure through the algorithm. Thus, the algorithm proposed in this paper was compared with the algorithm Lassen et al. (2013) by the average distance between the manual segmented and the automatically segmented fissure. The proposed algorithm with the thinning in the end achieved a shorter distance average value and a lower standard deviation compared with the method of (LASSEN et al., 2013). Finally, the results obtained for automatic segmentation of lung fissures are presented. The low incidence of false negative detections detection results, together with the significant reduction in false positive detections result in a high rate of settlement. We conclude that the segmentation technique for lung fissures is a useful target for pulmonary fissures on CT images and has potential to integrate systems that help medical diagnosis