A liver fibrosis staging method using cross-contrast network. (15th September 2019)

• Record Type: Journal Article
• Authors: Huang, Yudan; Chen, Ying; Zhu, Haochuan; Li, Weifeng; Ge, Yun; Huang, Xiaolin; He, Jian
• Is Part Of: Expert systems with applications. Volume 130(2019)
• Page Start: 124
• Page End: 131
• Publication Date: 2019-09-15
• Subjects: Convolutional neural network -- IBS theory -- Liver fibrosis -- Cross-contrast; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2019.03.049
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Mining complex texture features is critical for accurate liver fibrosis staging. It is the first experimental study combines IBS with convolutional network. Computational results suggest the network performs well on average. Abstract: In this paper we proposes a cross-contrast neural network (CCNN) for liver fibrosis classification. This method consists of two main parts. The first part extracts feature and gets the cross probability maps for utilizing the implicit contrast information among the inputs. The second part measures the similarity between two maps using the modified information based similarity (IBS) theory. IBS theory is a statistical method quantifies similarity between symbols and have been proved valid in many areas (Yang, Hseu, Yien, Goldberger, & Peng, 2003), but it has not been combined with neural network so far. CCNN combines the advantages of statistical analysis and convolutional neural networks, fitting the problem that the number of medical images is relatively small for traditional deep neural network to train. We apply CCNN on a 34-person dataset (23/11 for train/test) and the experimental results (shown in Table 3) clearly demonstrate the efficiency of the method. The highest accuracy is achieved on binary classification of F3 vs. F4, F0 vs.F3 and F1 vs. F3, which are 98.33%. The accuracy of no-to-moderate fibrosis (F0-2) vs. advanced fibrosis (F3-4) and 5 categories is 93.33% and 71.11% relatively. We find that most classification error occurs with F2. After removing F2, the classification accuracy of 4 categories rises to 84.44%.