(Tuesday, 7/31/2018) 3:45 PM - 4:15 PM
Room: Exhibit Hall | Forum 1
Purpose: The purpose of this study is to develop a deep learning-based computer-aided diagnosis (CAD)system for mammographic mass classification.
Methods: We developed a mass classification system that utilized three deep convolutional neural network (DCNN) architectures, named AlexNet, GoogLeNet, and VGG19 for benign and malignant classification of mass in mammograms. Experiments were conducted on a public mammogram database: Digital Database for Screening Mammography (DDSM).A transfer learning method was employed to training the DCNNs for mass classification. We first trained the DCNNs using about 1.3 million natural images for classification of 1000 categories. Then, we modified the output layer of each DCNN with two classes and subsequently trained the modified DCNN using 707 benign and 810 malignant mass images. The trained DCNNs were tested by using 241 mammographic images including 109 benign masses and 132 malignant masses.
Results: The experimental results showed that the sensitivity of AlexNet, GoogLeNet, and VGG19 were 72.7 %, 75.8 %, and 74.2 %, respectively, and the specificity of AlexNet, GoogLeNet, and VGG19 were 55.9 %, 60.0 %, and 56.8 %, respectively. The area under the receiver operating characteristic curves of AlexNet, GoogLeNet, and VGG19 were 0.74, 0.77, and 0.75, respectively.
Conclusion: This study presented the DCNN, which is one of the most important aspects in deep learning technique, for mammographic mass classification. Experimental results demonstrated that the DCNN has an excellent and a superior performance in mass classification in comparison with traditional mass classification methods. In addition, our study also demonstrated that the transfer learning is an effective training method for mammogram image modality even the scale of training data is relatively small.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
