The liver is an essential organ with numerous critical functions such as nutrient metabolism, detoxification and protein synthesis. Liver lesions can severely impair these functions, significantly impacting overall health. In particular, liver lesions often lack clear symptoms in the early stages, making early detection difficult. Without timely diagnosis, liver lesions can progress to serious complications, notably liver cancer, which was the third leading cause of cancer-related deaths worldwide in 2020. In this paper, we propose using various machine learning models on real-world datasets to detect and classify liver lesions, aiming to improve screening efficiency and early diagnosis. We have applied neural network models such as DenseNet-121, VGG-19 and ViT. Using a dataset comprising 2008 CT liver images divided into arterial, delay, plain and venous phases, we aimed to detect three types of lesions: liver cysts, hemangiomas and hepatocellular carcinoma. Experimental results show that ViT models achieved the highest accuracy of up to 99% with the shortest training. Manual data labeling requires a large workforce with high expertise and sometimes the quality of labeling is inconsistent due to differences between experts. Furthermore, as data becomes more complex and diverse, ensuring consistency and accuracy in the labeling process becomes increasingly challenging. To address these challenges, we propose an active learning solution. This solution can help automate the labeling process, saving time and cost while enhancing data quality. By effectively selecting data samples that need to be labeled, active learning can improve the accuracy and consistency of labeling.