Ultrasound Image-Based Liver Disease Detection Using Transfer Learning: A Comparative Analysis of CNN and Their Variants

Liver disease is a worldwide health concern, and accurate early detection is essential to prevent severe complications. Ultrasound imaging is frequently used for liver assessment because it is safe and affordable, but interpretation can be subjective and error prone. This study compares the performance of convolutional neural networks (CNNs)EfficientNet, MobileNet, and ConvNeXt variants in the publicly accessible annotated dataset of ultrasound images into benign, malignant, and normal categories. Results show that MobileNetV2 achieved the highest overall performance with an accuracy of 78%, recall of 78%, AUROC of 0.89 and inference time 28.47 MS, while EfficientNetB2 also displayed a strong balance between all metrics with an accuracy of 77% and inference time$\mathbf{7 4. 6 8}$MS. In contrast, ConvNeXt models offered competitive but not superior performance, managing an average accuracy of 74% and high inference time. These results suggest that lightweight and moderately scaled CNNs, which strike a balance between accuracy and computational efficiency, show promise for liver disease detection from ultrasound images, with MobileNetV2 benefiting from a reduced parameter count and energy-efficient design that enables faster inference and lower computational cost.