A comprehensive hierarchical classification based on multi-features of breast DCE-MRI for cancer diagnosis

Computer-aided diagnosis (CAD) is widely used for early diagnosis of breast cancer. The commonly used morphological feature (MF), dynamic feature (DF), and texture feature (TF) from breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) have been proved very valuable and are studied in this paper. However, previous studies ignored the prior knowledge that most of the benign lesions have clearer and smoother edges than malignant ones. Therefore, two new TFs are proposed. To obtain an optimal feature subset and an accurate classification result, feature selection is applied in this paper. Moreover, most existing CAD models with simple structure only focus on common lesions and ignore hard-to-spot lesions so that a satisfied performance can be obtained for common lesions but there are some contradictions for those hard-to-spot lesions. Therefore, in this paper, a comprehensive hierarchical model is proposed to deal with contradictions and predict all kinds of lesions. The experimental result shows that the new features obviously increase ACC of TF from 0.7788 to 0.8584 and feature selection increases ACC of DF form 0.6991 to 0.7345. More importantly, compared with the existing CAD models and deep learning method, the proposed model which provides a higher performance for both common and hard-to-spot lesions significantly increases the classification performance with sensitivity of 0.9452 and specificity of 0.9000.