Liver cancer is currently considered the most prevalent malignancy of the digestive system with a high mortality rate and unpredictable prognosis. Early detection and classification of liver lesions are crucial in formulating appropriate treatment strategies, ultimately extending patients’ survival time. This work proposes the utilization of three deep learning models - DenseNet-121, VGG-19, and Vision Transformer (ViT) - to detect and classify liver lesions using a real-world dataset comprising 2008 CT scans. The dataset is collected with four phases before and after contrast injection, and is used to identify three primary types of liver lesions: cysts, hemangiomas, and cellular carcinoma. To optimize the labor-intensive and costly labeling process, we propose an Active Learning approach to enhance the efficiency of semi-automated data annotation. Furthermore, this work integrates a Fuzzy Layer based on the Fuzzy C-Means algorithm into the ViT model, forming the Fuzzy-ViT framework. Additionally, the ViT and Fuzzy-ViT models are deployed in large-scale data environments, including Spark local and Spark cluster. Experimental results indicate that the ViT model is the most optimal choice due to its superior accuracy. The integration of Fuzzy C-Means significantly improves classification performance, achieving an accuracy of up to 98%. The implementation in a Spark cluster environment reduces training time by up to 50% compared to a local execution setup. These findings underscore the effectiveness of distributed computing in large-scale data processing. The study confirms that Fuzzy-ViT outperforms standard ViT, demonstrating the effectiveness of fuzzy logic in deep learning models for medical imaging. The research also highlights the trade-offs between accuracy and training speed in different computing environments, offering valuable insights for deploying AI-driven medical diagnostics at scale. • Developed a Vision Transformer enhanced with Fuzzy C-Means for liver lesion classification. • Achieved 98% accuracy on a real-world CT dataset of 2008 scans. • Reduced training time by 50% using distributed Spark cluster deployment. • Optimized annotation efficiency through Active Learning integration.
Phan et al. (Sun,) studied this question.