Parametric Image-based Breast Tumor Classification Using Convolutional Neural Network in the Contourlet Transform Domain

Automated visual identification of Benign and Malignant breast tumors even with ultrasound (US) B-Mode image is still an open area of research. This paper presents parametric image-based approach to classify and detect benign and malignant breast tumors from ultrasound images using a custom-made convolutional neural network architecture. The Rician Inverse Gaussian (RiIG) distribution is presented here as a suitable model for describing the statistics of the ultrasound images in the Contourlet Transform domain. Locally computed values of the dispersion parameters of the RiIG distribution in various contourlet sub-bands yield parametric images those are classified using the proposed convolutional neural network. Experiments are conducted on a publicly available dataset of 250 images, of 100 belong to the benign Fibroadenoma and 150 in the malignant category. The proposed method provides accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive values (NPV) of 96%, 97.3%, 94.12%, 96% and 96%, respectively. It is also shown that the accuracy obtained by the Proposed Method is higher than several recently reported results.