Seleção de atributos e classificação automática de lesões mamárias em imagens de ultrassom

Abstract

Breast cancer is one of the diseases that hit most women in the world. Due to the large number of factors associated with this type of disease, early detection is the best way to fight it. Mammography is the main imaging exam currently used for detection, since it is able to identify the presence of microcalcifications, which are a key indicator of the presence of cancer. As a complement exam, breast ultrasonography has also been widely used because of the large number of inconclusive mammograms and the difficulty of diagnosing younger women. However, the interpretation of ultrasound images is quite dependent on the experience of the doctor in charge of the diagnosis. To aid in the interpretation of these images, Computer-Aided Diagnosis (CAD) systems have appeared, and it seeks to provide a second opinion for medical specialists. In this work, the attribute selection and classification stages presented in these systems were developed. A wrapper approach with a search strategy based in genetic algorithms, and two filter approaches, the Welch's t test and the ReliefF algorithm was developed. To evaluate the subsets performance, a Multilayer Perceptron (MLP) neural network, with backpropagation learning algorithm was developed as a classifier. The metric used to evaluate the classification performance of each subset of attributes was the area of under the Receiver Operating Characteristic curve (Az).The used database has 541 images, with 314 benign lesions and 227 malignant lesions with a biopsyproven diagnosis. In addition, the database contains the manual segmentation of these images performed by a specialist physician and 22 morphological extracted attributes. The filter techniques results showed that some attributes alone are able to obtain good classification results, such as the depth/width ratio of the lesion, reaching 0.731 for Az. Besides that, the best results were found through the wrapper strategy, in which a value of 0.835 was obtained for Az using only eight of the 22 attributes, demonstrating the importance of these steps in this type of CAD system, increasing the final performance.