Cirrhosis classification based on MRI with duplicative-feature support vector machine (DFSVM)

•We propose a new CAD system for cirrhosis detection from normal hepatic tissue MR imaging.•We extract six texture features with different properties for charactering regions of interest.•Feature selection and classification are carried out with proposed DFSVM.•The impact of the most valuable features is well strengthened.•The high prediction performance can be got based on clinical data.

Magnetic resonance imaging (MRI) is a sensitive diagnostic method in improving the diagnostic capacity for hepatic cirrhosis and determining the accurate characterization of hepatic cirrhosis. But hepatic MRI has some shortcomings in detection and classification hepatic cirrhosis in clinical, especially using non-enhanced MRI for diagnosing early hepatic cirrhosis. And computer-aided diagnostic (CAD) system, including quantitative description of lesion and automatically classification, can provide radiologists or physicians an alternative second opinion to efficiently apply the abundant information of the hepatic MRI. However, it is expected to character comprehensively the lesion and guarantee high classification rate of CAD system. In this paper, a new CAD system for hepatic cirrhosis detection and classification from normal hepatic tissue non-enhanced MRI is presented. According to prior approach, six texture features with different properties based on gray level difference method are extracted from regions of interest (ROI). Then duplicative-feature support vector machine (DFSVM) is proposed for feature selection and classification: Firstly, the search process of DFSVM imitates diagnosis of doctors: doctor will take a more feature for consideration until the final diagnoses regardless of whether the feature is used in advance. So our algorithm is consistent with the process of clinical diagnosis. Secondly, the impact of the most valuable features will be well strengthened and then the high prediction performance can be got. Experimental results also illustrate the satisfying classification rate. Performance of extracted features and normalization are studied. And it is also compared with typical classifier ANN.