Encrypted login | home

Program Information

Transfer Learning for Mammogram Classification Using Pre-Trained Convolutional Neural Network

no image available
K Yasuda

K Yasuda1, H TSURU1 , M Ohki2 , (1) Graduate School of Medical Sciences, Kyushu University, Fukuoka, Fukuoka, (2) Graduate School of Medical Science, Kyushu University, Fukuoka, Fukuoka

Presentations

TU-C1-GePD-I-1 (Tuesday, August 1, 2017) 9:30 AM - 10:00 AM Room: Imaging ePoster Lounge


Purpose: This study was conducted to classify mammograms based on transfer learning using pre-trained convolutional neural network (CNN).

Methods: The subject materials in this study were 880 mammograms that were provided by the research group, Image Retrieval in Medical Applications. This image dataset contained five categories: no finding, benign calcification, malignant calcification, benign tumor, and malignant tumor. We chose 780 images randomly from the dataset as input to pre-trained CNN, and another 100 randomly selected images were used as the test dataset. For transfer learning, we used pre-trained CNN named Alexnet. This network is a deep CNN model trained on a large dataset of images. Transfer learning extracted the key image features from the mammograms and the multiclass support vector machine classified them. The prediction accuracies of each class were represented in a confusion matrix of the predicted and true classes.

Results: The prediction accuracies of no finding, benign calcification, malignant calcification, benign tumor, and malignant tumor were 70%, 60%, 35%, 65%, and 50%, respectively. The overall accuracy of the total classification was 56%.

Conclusion: Our transfer learning method classified the “no finding” class most accurately. Additionally, the discrimination between calcification and tumor was found to be adequate. Transfer learning using pre-trained CNN is a promising technique to assist the diagnosis in mammography.


Contact Email: