Uma abordagem para seleção de candidatos à região de lesão em imagens de mamografia digitalizadas utilizando algoritmos de detecção de pontos de interesse

Abstract

Breast cancer has been a growing problem for women all over the world. The correct interpretation of mamographic images is important for the diagnosis of breast cancer. However, this is a difficult task even for a specialist. Image processing is used to make the diagnosis less susceptible to errors. In this way, the present work proposes a new method for the selection of candidates for mass in mammography images, using algorithms to detect points of interest for the generation of the candidates. The behavior of SURF, SIFT, BRISK and ORB was analyzed. Furthermore, to prove the efficiency of the approach, the proposed method was compared with the algorithms EdgeBoxes, Rahtu, RandomPrim, Rantalankila, MCG and SelectiveSearch. We used 1210 mammography images acquired from the CBIS-DDSM database. Results showed that the proposed method, based on SURF algorithm presented the best performance, generating on average, for each image, 4.11 of considered candidates in the internal mass area and 10.20 candidates with threshold of IoU > 0.5 and reduction of exploratory space in 72%. Additionally, it was possible to obtain the best accuracy rate in location of masses, between the compared techniques, with 98.13% on average, best rate of medium accuracy with 0,34 mAP and lower execution cost, taking to only 0.001 seconds to generate each mass candidate.