An automatic breast computer-aided diagnosis scheme based on a weighted fusion of relevant features and a deep CNN classifier

Abstract

Mammography continues to play a central part in the early breast masses diagnosis and has raised several challenges in breast cancer detection. Nevertheless, it remains still difficult to detect abnormalities in a dense breast and some benign lesions may have similar appearance with masses. This study proposes a novel Computer-Aided Diagnosis (CADx) allowing benign and malignant mass classification. Accordingly, relevant textural-shape descriptors are proposed, which is the aggregation of a new Local Binary based feature, namely the Monogenic Gray Level and Local Difference (MGLLD), where both texture characteristics and breast tissue density information, and the Zernike moments are incorporated. A heuristic algorithm is then devised for optimizing the proposed features with respect to inter-class and intra-class distribution, by a convenient ponderation. Then, a set of classifiers are applied for opting to the best decision making solution. The Deep CNN yields the best accuracy, with an Area Under Curve (AUC) of 0.98 on Curated Breast Imaging Subset of DDSM (CBIS-DDSM). In addition, the authors are based on a subjective approval of the extracted Region Of Interest (ROI) by experts in radiology. The proposed scheme proves its effectiveness especially on some challenging breast cancer cases, corresponding to higher breast tissue density. In medical image processing, these results open new perspectives with respect to the use of the deep learning solutions, where the small sample-number is known to be a limitation (because pathologies cannot be intentionally provoked).