Abstract:Objective: This research aimed to improve diagnosis of non-alcoholic fatty liver disease (NAFLD) by deep learning with ultrasound Images and reduce the impact of the professional competence and personal bias of the diagnostician. Method: Three convolutional neural network models were used to classify and identify the ultrasound images to obtain the best network. Then, the features in the ultrasound images were extracted and a new convolutional neural network was created based on the best network. Finally, the accuracy of several networks was compared and the best network was evaluated using AUC. Results: Models of VGG16, ResNet50, and Inception-v3 were individually applied to classify and identify 710 ultrasound images containing NAFLD, demonstrating accuracies of 66.2%, 58.5%, and 59.2%, respectively. To further improve the classification accuracy, two features are presented: the ultrasound echo attenuation coefficient (θ), derived from fitting brightness values within sliding region of interest (ROIs), and the ratio of Doppler effect (ROD), identified through analyzing spots exhibiting the Doppler effect. Then, a multi-input deep learning network framework based on the VGG16 model is established, where the VGG16 model processes ultrasound image, while the fully connected layers handle θ and ROD. Ultimately, these components are combined to jointly generate predictions, demonstrating robust diagnostic capabilities for moderate to severe fatty liver (AUC = 0.95). Moreover, the average accuracy is increased from 64.8% to 77.5%, attributed to the introduction of two advanced features with domain knowledge. Conclusion: This research holds significant potential in aiding doctors for more precise and efficient diagnosis of ultrasound images related to NAFLD.

Method of edge energy template for detection of ultrasonic based fatty liver

Computer Aided Diagnosis of Fatty Liver Ultrasonic Images Based on Support Vector Machine

Combined Density, Texture and Shape Features of Multi-phase Contrast-Enhanced CT Images for CBIR of Focal Liver Lesions: A Preliminary Study

Computer-aided diagnosis of fatty liver based on ultrasonic images]

B-Scan Ultrasound Image Feature Extraction for Quantitative Grading of Fatty Liver Severity

Feature Analysis and Automatic Classification of B-mode Ultrasound Images of Fatty Liver

Fibrosis quantitative analysis based on ultrasound liver Image

Ultrasound Image Edge Detection Based on a Novel Multiplicative Gradient and Canny Operator.

Quantitative Analysis of Fatty Liver Using Gray-Scale Contrast-Enhanced Ultrasound

Quantitative analysis and identification of liver B-scan ultrasonic image based on BP neural network

The Effects of TIS and MI on the Texture Features in Ultrasonic Fatty Liver Images

A Retrieval System For 3d Multi-Phase Contrast-Enhanced Ct Images Of Focal Liver Lesions Based On Combined Bags Of Visual Words And Texture Words

Bio-Inspired Network for Diagnosing Liver Steatosis in Ultrasound Images

Multiparametric quantification and visualization of liver fat using ultrasound

Shape Detection of Liver From 2D Ultrasound Images

Automatic Classification Of Focal Liver Lesion In Ultrasound Images Based On Sparse Representation

[A Novel Method to Quantitatively Analyze the Fatty Liver Ultrasonic Images Based on Multiresolution Fractal Brownian Motion Model and Genetic Algorithm].

Deep Learning With Ultrasound Images Enhance the Diagnosis of Nonalcoholic Fatty Liver

Liver Fibrosis Identification Based on Ultrasound Images Captured under Varied Imaging Protocols

Extraction of Liver Capsule from High-Frequency Ultrasound Images Via Drift Iterative Search Algorithm

Research on Fuzzy Enhancement in the Diagnosis of Liver Tumor from B-mode Ultrasound Images