| Abstrakt: |
Breast cancer, if diagnosed in its early stages, can be treated more effectively. For this purpose, computer-aided diagnosis (CAD) systems based on ultrasound images have been proposed in order to detect breast lesions and discriminate between the benign and malignant types of these lesions. In this paper, a novel CAD system based on gated convolutional neural networks (CNN) is proposed, which can help classify breast lesions in ultrasound images and differentiate benign lesions from malignant ones. Specifically, the required feature maps are extracted by using a pre-trained CNN. A common strategy for improving the performance of the classification methods for breast lesions is to combine the feature maps learned at different layers; yet, most of the current methods combine these feature maps without adherence to feature map selection. The problem is that lack of feature map selection may impose redundant information on the network and may cause over-segmentation when more information from the lower layers is required. To tackle this problem, the present study proposes a gate mechanism for use with entropy maps in order to select the features. Subsequently, the selected feature maps are adaptively weighted according to their relative significance. According to the experimental results, the proposed method would significantly improve the accuracy, sensitivity, specificity and F1 score area under the ROC curve of the classification process, and the method would outperform the other breast cancer CAD systems. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink