Hybrid classification of diffuse liver diseases in ultrasound images using deep convolutional neural networks

• DOI: 10.1016/j.imu.2020.100496
• Publication Date: 2020-12-11T01:47:48Z
• Authors: Pezhman Pasyar; Tahereh Mahmoudi; Seyedeh-Zahra Mou...; Alireza Ahmadian; Hossein Arabalibeik; Nahid Soltanian; Amir Reza Radmard
• Keywords: Liver disease; Majority Rule

Despite the fact that liver biopsy is considered to be gold standard for detecting diffuse diseases, it an invasive method with numerous side effects. Diffuse diagnosis using ultrasound imaging may influenced by Physician subjectivity. Therefore, accurate classification of diseases remains a notable demand. In this study, categorize status, novel deep classifier, comprised pre-trained convolutional neural networks (CNNs) proposed. Several networks, namely ResNeXt, ResNet18, ResNet34, ResNet50, and AlexNet which concatenated fully connected (FCNs) are used. Extracted features transfer learning can provide sufficient information. An FCN then put images into different states disease, normal liver, hepatitis, cirrhosis. Two-class (normal/cirrhosis, normal/hepatitis, cirrhosis/hepatitis) three-class (normal/cirrhosis/hepatitis) classifiers were trained distinguish these images. Since two-class showed better performance compared classifiers, hybrid classifier proposed so as integrate weighted probabilities classes obtained means each individual classifier. Then, majority voting strategy employed select class higher score. The experimental results show accuracy 86.4% ResNet50 classified three classes. distinction between cirrhosis well hepatitis demonstrate sensitivity specificity first group 90.9% latter shows 90.9%, 81.8%.