Mammogram Screening for Breast Density Classification using a soft voting ensemble of Swin Transformers and ConvNext models
Muhammad Hussain, Fahman Saeed, Mariam Busaleh, Hatim Aboalsamh, Fazal-e-Amin · 2022
Breast cancer risk is increased by dense breast tissue. The existing mammographic density classification methods cannot provide sufficient classification accuracy. The classification of breast density is still a challenging issue. In this paper, we propose an ensemble method for improving the accuracy of mammographic breast density classification. The ensemble method uses two views of screening mammography and employs state-of-the-art deep learning-based methods such as Swin Transformer and ConvNeXts model as base learners for local decisions and soft voting to predict the final decision. The method has been thoroughly evaluated on the benchmark DDSM dataset; it achieved an accuracy of 97.61%, a sensitivity of 95.11%, and a specificity of 98.32%. The comparison shows that the method significantly outperforms the state-of-the-art methods for breast density classification. This study provides insight into the application of deep learning to classify breast density.