Feature selection by genetic and swarm algorithms for classifying breast ultrasound images

Designing a Computer Aided Diagnosis (CAD) system for Breast Ultrasound (BUS) B-mode image is a crucial task in medical diagnosis. Selecting and finding an optimal Feature Subset to classify the Benign and Malignant tumor more accurately is a high dominant process. This work emphasizes on analyzing and identifying the most discriminating features for diagnosing breast cancer using BUS images. In this work, 71 features (28 Morphological, 43 Textural) are extracted from 84 images (49 Benign, 35 Malignant cases) and considered for Feature Selection process. As a feature selector, Principal Feature Analysis (PFA) is applied to reduce the dimensionality of large feature vector by selecting 35 uncorrelated features. Finally GA, PSO, ACO and ABC optimization algorithms wrapped with Artificial Neural Network (ANN) as fitness evaluator are designed to mine the optimal feature subsets from the 35 selected features of PFA. Three feature subsets each containing 15, 10 and 5 features obtained from each optimization technique are assessed using the metrics like Sensitivity, Specificity, Accuracy, Matthews Correlation Coefficient (MCC) and Az under Receiver Operating Characteristics (ROC) curve. The results show that the PSO achieved a better classification accuracy of 94% f-measure, 0.8767 MCC value and Az as 0.9334 for all feature subsets than Bio and Swarm algorithms whereas the 12 common features selected from GA and Swarm algorithms yields 97.62% classification accuracy and suits wellfor designing a Breast Ultrasound CAD system.