Ultrasound-based liver disease classification using deep convolutional neural networks

Abstract

Liver disease is a significant health concern in the world, which has led to millions of deaths annually. Complications associated with the disease include cirrhosis, liver failure, and hepatocellular carcinoma. Early and correct diagnosis minimizes these risks, enhances treatment results and facilitates timely clinical decisions. One of the most common diagnostic methods to assess the liver is ultrasound imaging, as it is safe, non-invasive, cheap and can be used in both urban and resource-restricted healthcare facilities. Nonetheless, ultrasound interpretation is usually subjective and relies on operator expertise, level of experience and changes in image quality despite its availability. This renders automated and standardized diagnostic support as a significant field of study. This paper explores the potential of modern convolutional neural network (CNN) architectures to aid in classifying liver ultrasound images into three clinically relevant categories, which include normal, benign, and malignant. We concentrate on three popular CNN families EfficientNet, MobileNet, and ConvNeXt each with different design philosophies in deep learning. MobileNet focuses on lightweight operations and algorithms, EfficientNet provides balanced scaling of compounds to enhance accuracy at fewer parameters, and ConvNeXt applies modernized convolutional architectures based on the transformer-like principles. Each of the models was trained in a frozen-backbone transfer learning configuration, with all experiments being fairly compared by maintaining similar preprocessing, training parameters, and classification layers. This study uses a dataset of annotated ultrasound images, which is publicly available. Though the dataset is quite small and unbalanced, it is representative of real-world clinical situations in which malignant cases are more likely to be reported. All the models were trained on a single pipeline. To measure performance, Accuracy, Precision, Recall, F1- Score, and AUROC were used. MobileNetV2 was the most successful of all the tested models, with an F1-Score of 75.20%, an Accuracy of 78.37%, and a recall of 78.03%. These findings indicate that not only did MobileNetV2 make accurate predictions but also it was highly effective in the detection of malignant cases. EfficientNetB2 was also quite stable, achieving 77.02% Accuracy and showing good balance in all measures. ConvNeXt variants were competitive but did not outperform MobileNetV2 or EfficientNetB2, probably because they have more capacity and are sensitive to small datasets trained with frozen backbone. Overall, the findings show that lightweight and moderately sized CNNs can provide strong diagnostic support for liver ultrasound classification without requiring large number of computational resources. These results highlight the potential for accessible AI-driven tools that support radiologists, reduce interpretation variability, and improve early detection of liver disease.