Grading of steatosis, fibrosis, lobular inflammation, and ballooning from liver pathology images using pre-trained convolutional neural networks

This study aims to automatically detect the degree of pathological indices as a reference method for detecting the severity and extent of various liver diseases from pathological images of liver tissue with the help of deep learning algorithms. Grading is done using a collection of pre‐trained convolutional neural networks, including DenseNet121, ResNet50, inceptionv3, MobileNet, EfficientNet‐b1, EfficientNet‐b4, Xception, NASNetMobile, and Vgg16. These algorithms are performed by fine‐tuning the trainable layers of the networks. The results showed that compared to other methods, the EfficientNet‐b1 network provides a better response to grade the stage of liver disease among all indicators from pathological images, due to its structural features. This classification accuracy was 97.26% for fibrosis, 94.1% for steatosis, 90.2% for lobular inflammation, and 98.0% for ballooning. Consequently, this fully automated framework can be very useful in clinical methods and be considered as an assistant or an alternative to the diagnosis of experienced pathologists.