A two-stage multi-view learning framework based computer-aided diagnosis of liver tumors with contrast enhanced ultrasound images

Article type: Research Article

Authors: Guo, Le-Hang^{a; 1} | Wang, Dan^{a; 1} | Qian, Yi-Yi^b | Zheng, Xiao^b | Zhao, Chong-Ke^a | Li, Xiao-Long^a | Bo, Xiao-Wan^a | Yue, Wen-Wen^a | Zhang, Qi^b | Shi, Jun^{b; *} | Xu, Hui-Xiong^{a; *}

Affiliations: [a] Department of Medical Ultrasound, Shanghai Tenth People’s Hospital, Ultrasound Research and Education Institute, Tongji University School of Medicine, Shanghai, China | [b] Shanghai Institute for Advanced Communication and Data Science, School of Communication and Information Engineering, Shanghai University, Shanghai, China

Correspondence: [*] Corresponding author: Hui-Xiong Xu, MD, PhD, Department of Medical Ultrasound, Shanghai Tenth People’s Hospital, Ultrasound Research and Education Institute, Tongji University School of Medicine, Shanghai 200072, China. E-mail: xuhuixiong@126.com and Jun Shi, PhD, Shanghai Institute for Advanced Communication and Data Science, School of Communication and Information Engineering, Shanghai University, Shanghai, 200444, China.

Note: [1] These authors contributed equally to this work.

Abstract: OBJECTIVE:With the fast development of artificial intelligence techniques, we proposed a novel two-stage multi-view learning framework for the contrast-enhanced ultrasound (CEUS) based computer-aided diagnosis for liver tumors, which adopted only three typical CEUS images selected from the arterial phase, portal venous phase and late phase. MATERIALS AND METHODS:In the first stage, the deep canonical correlation analysis (DCCA) was performed on three image pairs between the arterial and portal venous phases, arterial and delayed phases, and portal venous and delayed phases respectively, which then generated total six-view features. While in the second stage, these multi-view features were then fed to a multiple kernel learning (MKL) based classifier to further promote the diagnosis result. Two MKL classification algorithms were evaluated in this MKL-based classification framework. We evaluated proposed DCCA-MKL framework on 93 lesions (47 malignant cancers vs. 46 benign tumors). RESULTS:The proposed DCCA-MKL framework achieved the mean classification accuracy, sensitivity, specificity, Youden index, false positive rate, and false negative rate of 90.41 ± 5.80%, 93.56 ± 5.90%, 86.89 ± 9.38%, 79.44 ± 11.83%, 13.11 ± 9.38% and 6.44 ± 5.90%, respectively, by soft margin MKL classifier. CONCLUSION:The experimental results indicate that the proposed DCCA-MKL framework achieves best performance for discriminating benign liver tumors from malignant liver cancers. Moreover, it is also proved that the three-phase CEUS image based CAD is feasible for liver tumors with the proposed DCCA-MKL framework.

Keywords: Contrast-enhanced ultrasound, artificial intelligence, liver tumor, deep canonical correlation analysis, multiple kernel learning

DOI: 10.3233/CH-170275

Journal: Clinical Hemorheology and Microcirculation, vol. 69, no. 3, pp. 343-354, 2018