BI-RADS classification of breast masses based on deep contourlet features

Breast cancer is highly dominant in women and is also a leading cause of death. The high mortality rate is contributed mainly due to late diagnosis which can be reduced by manifolds if detected and diagnosed at the earliest. In recent years, there has been a significant shift from traditional machine learning techniques to deep learning approaches across various fields, including object detection, image recognition, and biomedical applications. In this paper, a new deep learning architecture that integrates the contourlet transform with the DenseNet-121 framework is proposed for the classification of masses according to the BI-RADS categories (2, 3, 4, and 5) as per Breast Imaging Reporting and Data System as well as into benign and malignant classes on digital mammograms. DenseNet demonstrates a direct connection between any two layers that share the same feature map size. The Contourlet transform extends wavelet transforms into two dimensions by incorporating multiscale and directional filter banks which makes it particularly suitable for applications of capturing edge and contour information in images. The mass lesions are detected via U-Net architecture and features are automatically learned through the DenseNet-121 model in combination with contourlet transform. The assessment is conducted on INbreast and Digital Database of Screening Mammography datasets are a collection of mammogram images used extensively in research and development related to breast cancer analysis, which are publicly available datasets. The performance of the proposed model is promising when compared with the related schemes in the state-of-the-art.