Atualização local automática de pesos para recuperação de nódulos similares de câncer pulmonar

Abstract

Lung cancer has become the most lethal malignancy in the world in recent decades. And despite advances in medicine, there has been little progress regarding the cure of the disease. According to the National Cancer Institute in the last global estimate of the incidence of lung cancer in 2012, there were 1.82 million cases of cancer, with 1.24 million among men and 583 thusand among women. The main cause of lung cancer is smoking that is responsible for 90 % of diagnosed cases. The diagnosis of lung cancer is done mainly based on CT images, and today it is considered the main visualization technique for detecting pulmonary nodules. However, the process of identifying and classi cation of nodules are complex and involves subjective and qualitative factors that lead experts to error. This scenario requires the use of computational techniques to e ectively manipulate the data and provide the means for more accurate diagnoses. Computer systems have been developed in order to search and retrieve imaging exams already diagnosed which are similar to a new case with unknown pathology according to the similarity between their characteristics. This property is intrinsic to Content-Based Image Retrieval (CBIR). Diagnosed exams retrieved can be used as a second opinion to guide those specialists in the diagnosis, providing more information. However, CBIR presents some limitations regarding to the process of segmentation and representation of image characteristics through of attributes, as well as determine an appropriate similarity metric. This paper presents a local update weighing algorithm applied to the Weighted Euclidean Distance (WED) in a CBIR architecture in order to verify if the WED with adjusted weights is more accurate than the Euclidean Distance in image retrieval of pulmonary nodules. For this, the 3D Texture Attributes (3D AT) and 3D Margin Sharpness Attributes (3D MSA) were used to represent nodules. Presente process consists of two phases that are performed sequentially and cyclically being an Assessment Phase and Training Phase. At each iteration the weights are adjusted according to the retrieved nodules. At the end of cycles execution, it is obtained a set of attribute weights that optimize the recovery of similar nodes. The results achieved by updating the weights were promising and increase precision by 10% to 6% on mean for recovery of benign and malignant nodules respectively with recall 25%. In the best case, the 3D MSA provided 100% of precision for the two classes with recall 90%. This proves the e ectiveness of the algorithm achieving the goals to this work and con rms the hypothesis that the DEP, with adjusted weights, provides greater precision than DE as a similarity metric in CBIR systems.