A classification framework for prediction of breast density using an ensemble of neural network classifiers

• Published in: Biocybernetics and biomedical engineering 2017, Vol.37 (1), p.217-228
• Main Authors: Kumar, Indrajeet, Bhadauria, H.S, Virmani, Jitendra, Thakur, Shruti
• Format: Article
• Language: English
• Subjects: Advanced Basic Science; Breast density classification; Ensemble classifier; Gray level co-occurrence matrix; Internal Medicine; Mammography; Neural network classifier
• ISSN: 0208-5216
• DOI: 10.1016/j.bbe.2017.01.001

Abstract The present work proposes a classification framework for the prediction of breast density using an ensemble of neural network classifiers. Expert radiologists, visualize the textural characteristics of center region of a breast to distinguish between different breast density classes. Accordingly, ROIs of fixed size are cropped from the center location of the breast tissue and GLCM mean features are computed for each ROI by varying inter-pixel distance ‘ d’ from 1 to 15. The proposed classification framework consists of two stages, ( a ) first stage : this stage consists of a single 4-class neural network classifier NN0 (B-I/B-II/B-III/B-IV) which yields the output probability vector [ PB-I PB-II PB-III PB-IV ] indicating the probability values with which a test ROI belongs to a particular breast density class. ( b ) second stage : this stage consists of an ensemble of six binary neural network classifiers NN1 (B-I/B-II), NN2 (B-I/B-III), NN3 (B-I/B-IV), NN4 (B-II/B-III), NN5 (B-II/B-IV) and NN6 (B-III/B-IV). The output of the first stage of the classification framework, i.e. output on NN0 is used to obtain the two most probable classes for a test ROI. In the second stage this test ROI is passed through one of the binary neural networks, i.e. NN1 to NN6 corresponding to the two most probable classes predicted by NN0. After passing the entire test ROIs through the second stage, the overall accuracy increases from 79.5% to 90.8%. The promising results achieved by the proposed classification framework indicate that it can be used in clinical environment for differentiation between breast density patterns.