Journal
JOURNAL OF COMPUTATIONAL SCIENCE
Volume 27, Issue -, Pages 57-68Publisher
ELSEVIER
DOI: 10.1016/j.jocs.2018.05.005
Keywords
Breast cancer; Rank-based stochastic pooling; Parametric rectified linear unit; Convolutional neural network; Cost-sensitive learning
Funding
- Natural Science Foundation of China [61602250]
- Natural Science Foundation of Jiangsu Province [BK20150983]
- Program of Natural Science Research of Jiangsu Higher Education Institutions [16KJB520025]
- Open fund for Jiangsu Key Laboratory of Advanced Manufacturing Technology [HGAMTL1601, HGAMTL-1703]
- Open fund of Key Laboratory of Guangxi High Schools Complex System and Computational Intelligence [2016CSC101]
- Open Fund of Guangxi Key Laboratory of Manufacturing System & Advanced Manufacturing Technology [17-259-05-011K]
- Henan Key Research and Development Project [182102310629]
Ask authors/readers for more resources
Aim: Abnormal breast appears similar as dense breast in mammography, which makes it a challenge for radiologists to identify. Scholars have proposed numerous computer-vision and machine-learning based approaches. Nevertheless, the features were manually designed. Method: In this study, the breast dataset was chosen as the open-access mini MIAS dataset. Cost-sensitive learning was used to balance the dataset. Data augmentation was used to increase the size of training set. We proposed an improved nine-layer convolutional neural network (CNN). In addition, we compared three activation functions: rectified linear unit (ReLU), leaky ReLU, and parametric ReLU. Besides, six pooling techniques were compared: average pooling, max pooling, stochastic pooling, rank-based average pooling, rank-based weighted pooling, and rank-based stochastic pooling. Results: The results over 100 test set showed the combination of parametric ReLU and rank-based stochastic pooling performed the best, with sensitivity of 93.4%, specificity of 94.6%, precision of 94.5%, and accuracy of 94.0%. This result is better than six state-of-the-art breast cancer detection approaches. Conclusion: Deep learning can provide better detection results than traditional artificial intelligence methods. We validate why we set the number of convolution layers as 2. We shall try to further improve the performance of this proposed method. Crown Copyright (C) 2018 Published by Elsevier B.V. All rights reserved.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available