Classification of Nonalcoholic Fatty Liver Grades using Pre-Trained Convolutional Neural Networks and a Random Forest Classifier on B-Mode Ultrasound Images.

Background: Nonalcoholic Fatty Liver Disease (NAFLD) as a prevalent condition can significantly have health implications. Early detection and accurate grading of NAFLD are essential for effective management and treatment of the disease. Objective: The current study aimed to develop an advanced hybrid machine-learning model to classify NAFLD grades using ultrasound images. Material and Methods: In this analytical study, ultrasound images were obtained from 55 highly obese individuals, who had undergone bariatric surgery and used histological results from liver biopsies as a reference for NAFLD grading. The features were extracted from the ultrasound images using popular pretrained Convolutional Neural Network (CNN) models, including VGG19, MobileNet, Xception, Inception-V3, ResNet-101, DenseNet-121, and EfficientNet-B7. The fully connected layers were removed from the CNN models and also used the remaining structure as a feature extractor. The most relevant features were then selected using the minimum Redundancy Maximum Relevance (mRMR) method. We then used four classification algorithms: Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Multilayer Perceptron (MLP) neural network, and Random Forest (RF) classifiers, to categorize the ultrasound images into four groups based on liver fat level (healthy liver, low fat liver, moderate fat liver, and high-fat liver). Results: Among the different CNN models and classification methods, EfficientNet-B7 and RF achieved the highest accuracy. The average accuracies of the LDA, MLP, SVM, and RF classifiers for the feature extraction method with EfficientNet-B7 were 88.48%, 93.15%, 95.47%, and 96.83%, respectively. The proposed automatic model can classify NAFLD grades with a remarkable accuracy of 96.83%. Conclusion: The proposed automatic classification model using EfficientNet-B7 for feature extraction and a Random Forest classifier can improve NAFLD diagnosis, especially in regions, in which access to professional and experienced medical experts is limited.