چكيده به لاتين
Breast cancer can be considered as the most common cancer and also the second leading cause of death by cancer among women in the world. Hence, finding ways to diagnose and treat this disease is a critical and sensitive challenge in health of human community. Various methods are proposed for breast screening in women, and one of the non-invasive of these methods is magnetic resonance imaging. This method is one of the most commonly used methods of breast cancer diagnosis today. However, lesions do not have borderline characteristics and morphologic features of their own. Therefore, differentiating between benign and malignant lesions in a normal state of work is very time consuming and challenging. In this study, a computer-aided auto-diagnosis system is developed for the diagnosis and classification of axial magnetic resonance images of breasts in two classes of benign and malignant.
Initially, suspected parts of the lesion are separated as a rectangular box around the lesion by an experienced radiologist. Then a proposed algorithm is used to precisely separate the lesion. The proposed separation algorithm segmentation the lesions considering unevenness of the images for the first time and attempts to remove the false positive regions using morphological operations and removing vein. In the next stage, four groups of features are extracted from the separated parts of the lesions where each of them expresses particular states of the lesion structure.
These four groups are tissue features, synthetic, frequency, and morphology. Here a new group of features is extracted called the Gabor-Haralik features, which presents a particular efficiency. Due to high dimensions of the features, different methods of selecting features have been conducted in this research based on the optimization algorithms. The HMSFLA optimization algorithm is presented here for this purpose and is used for the training of MLP-NN and ANFIS after selecting the most appropriate features in the classification stage. In all these stages, 46 lesions are used, and their sensitivity, specificity, accuracy, and F measure have been reported as 90, 96/1, 93/483 and 92/30%. These values indicate the efficiency of the proposed diagnosis system in the classification of benign and malignant lesions in magnetic resonance imaging of the breast.
