Artificial neural network to aid differentiation of malignant and benign breast masses by ultrasound imaging

Abstract

The goal of this study is to evaluate an Artificial Neural Network (ANN) for differentiating benign and malignant breast masses on ultrasound scans. The ANN was designed with three layers (input, hidden and output layer), where a sigmoidal (hyperbolic tangent) response function is used as an activation function at each unit. Data from 54 patients with biopsy-proven malignant (N=20) and benign (N=34) masses were used to evaluate the diagnostic performance of the ANN. Of the seven quantitative features extracted from ultrasound images, only four showed statistically significant difference between the two categories. These features were margin sharpness, margin echogenicity, angular continuity, and age of patients. The diagnostic performance was evaluated by round-robin substitution to negate bias due to small sample size. All the input features were standardized to zero-mean and unit-variance to prevent non-uniform learning, which can generate unwanted error. The outputs of the network were analyzed by Receiver Operating Characteristics (ROC). The resulting area under the ROC curve Az was 0.856 ± 0.058 with 95% confidence limit from 0.734 to 0.936, providing 76.5% specificity at 95% sensitivity. The performance of the ANN was comparable to the performance by logistic regression analysis reported by our group earlier. These results suggest that an ANN when combined with sonography can effectively classify malignant and benign breast lesions.