Automatic Recognition of Non-alcoholic Fatty Liver by using Transfer Deep Learning Networks from Ultrasound Images

Hamed Zamanian, Ahmad Shalbaf



One of the common diseases is fatty liver disease, which can be seen significantly in patients with excessive obesity as well as patients with diabetes. Using ultrasound imaging methods, it can be possible to identify and evaluate patients affected by this disease. The aim of this study was to use advanced machine learning algorithms for better and more accurate classification of the acquired images from different patients affected by fatty liver disease.

Materials and Methods:

 In this study, the acquired ultrasound images of 55 patients suspicious as having fatty liver disease have been used. The level of fat for different patients was labeled by biopsy sampling. Based on this experiment, the patients were predicated as having a fatty liver when more than 5% of their liver hepatocytes were accompanied by infiltration of fat. Then, we utilized some pre-trained convolutional neural networks, including Inception-ResNetV2, GoogleNet, and AlexNet to extract high-level features of the collection of the acquired images. After that, a SoftMax layer was implemented to classify the images that indicated fatty liver.


 The resulted precisions for Inception-ResNetV2, GoogleNet, and AlexNet pre-trained convolutional networks were 0.8108, 0.9459, and 0.9932, respectively. Also, the area under the curve of receiver operating characteristics (AUC) for these networks were 0.9757, 0.9960, and 0.9963, respectively.


The proposed intelligent algorithm can help sonography experts to recognize the liver tissues with fat automatically and accurately without the need for a specialist to assign the region of interest for evaluation.


Fatty liver, Ultrasound imaging, Machine learning, Deep transfer learning networks

Full Text:


Copyright (c) 2020 GOVARESH

Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.