Breast Mass Recognition Model via Deep-Level Pathological Information Mining

Abstract

Breast cancer is the most common variant of cancer in women. Breast mass recognition model can assist pathologists in formulating their clinical diagnoses more efficiently. However, sample scarcity in the field of medical image analysis usually causes the overfitting problem. A novel breast mass recognition model via deep-level pathological information mining is proposed in this paper to address this problem. Firstly, a new sample refinement strategy is constructed to obtain image samples with high-quality across different mammographic datasets, which mainly deals with the sample scarcity problem from the data augmentation perspective. The deep-level pathological information contained in the limited labeled samples is mined out in turn from the shallower to the deeper, which mainly copes with the sample scarcity problem from the feature selection perspective. A novel feature selection algorithm called multi-view efficient range-based gene selection (MvERGS) is proposed to improve the discriminant ability of each image feature and reduce the corresponding dimensions, which helps to fit sample size well. Then the state-of-the-art discriminant correlation analysis (DCA) method is employed to analyze the deep cross-modal correlations among diverse refined features, which is used to depict the lesion areas in mammographs more accurately. Finally, based on deep-level pathological information and traditional classifier, an effective breast mass classification model is trained. Extensive experimental results demonstrate that the proposed breast mass classification model is superior to most baselines in some key metrics, including Accuracy and AUC, and it can cope with the overfitting problem very well.

Keywords

• |breast mass recognition|pathological information mining|sample refinement|feature selection|multi-view efficient range-based gene selection (mvergs)