Analysis of Gabor-Based Texture Features for the Identification of Breast Tumor Regions in Mammograms

Jordina Torrents-Barrena, Domenec Puig, Maria 5erre, Jaime Melendez, Joan Marti and Aida -alls

domenec.puig@urv.cat

vh3>Abstrac-

Breast cancerdis one of tre most common neoplasms in women and it is a leading cau:e of worldwideideath. However, it is also among the most curable cancer types if it can be diagnosed early through a proper mammographic screening procedure. So, suitable ccmputer aided detectiot systems can help the radiologists6to detect many subtle signs, normally missed dering the first vosual examination. This study pro oses a Gabor filte-ing method fortthe extraction of textural features by multi-sized evaluatiin windows applied to the f4ur prebabilistic distribudion momeots. Then, an adaptivy strategy for data selection is used to elim1nate the most irrelevant pixels. Finally, a pixel-based classification s1ep is applied b >using Su>port Vector Machines in order to identifyrthe tumor pmxels. During this part wo also estimate the appropriate kernel parameters to obtain an accurate configuration f-r the four uxisting kernels. Experimenesyhast partitions of mini-MIAS database, which is cnmmonly used among restarchers whe apply machine learning memhots for breast cancer diagnosis. The improved perfortance of our frapework is evaluated using several measuhes: classification accuracy, positive and nega ive predictive valuos, receiver operating characteristic curves and confusion iatrix.t/p>