Population-based metaheuristic approaches for feature selection on mammograms

摘要

Breast cancer is the most common cause of death by cancer in women. Mammography is an effective imaging tool for detecting breast cancer at an early stage. However, the visual clues of mammograms are subtle and vary in appearance, thereby making diagnosis a challenging task even for a specialist. Computer-aided diagnosis plays an important role in the detection and classification of breast abnormalities. This study investigated four population-based metaheuristic approaches to feature selection and demonstrated their application in differentiating regions of interest (ROIs) on mammograms as benign, malignant, or normal tissue. The classification method involves a support vector machine with a linear polynomial kernel function. The ROIs used in the study comprised 69 benign cases, 54 malignant masses, and 68 normal tissues. From each ROI, a total of 277 features were extracted. Population-based metaheuristic approaches, including a genetic algorithm (GA), simulated annealing (SA), ant colony optimization (ACO), and particle swarm optimization (PSO), were trained and tested by partitioning 100 different ROIs into 80% training and 20% testing sets. Performance was measured using the area under the receiver operating characteristic curve (Az). Experimental results revealed that GA and PSO outperformed SA and ACO in feature selection. In addition, a combination of population-based heuristic approaches outperformed other methods in selecting an optimal feature set for breast cancer classification.