机构
放射科
医疗保健
计算机科学
医学
医学物理学
政治学
法学
作者
Chiu‐Han Hsiao,Tzu-Lung Sun,Yen-Yen Liao,Yeong-Sung Frank Lin,Chih‐Horng Wu,Yu-Chun Lai,Hung-Pei Wu,Pin-Ruei Liu,Bo-Ren Xiao,Chien‐Hung Chen,Yennun Huang
摘要
Hepatocellular carcinoma, a form of liver cancer, is the most common liver cancer. It can be detected noninvasively through medical imaging, such as computed tomography (CT), which is the most common diagnostic tool for liver tumors. However, the interpretation of CT results requires considerable expertise and professional knowledge, potentially risking misjudgment. This paper introduces a method that employs a hybrid of two- and three-dimensional deep learning models and a federated learning framework to identify liver and tumor regions in medical images. Two models were trained: one for identifying liver regions and another for determining tumor areas. After we leveraged the first model's output as input data for the second model, noise from nonliver parts was reduced. In addition, the federated learning framework increased data set diversity and performance by aggregating deep-learning model parameters from various hospitals. Our training data comprised 131 CT scans from the MICCAI 2017 Liver and Liver Tumor Segmentation Challenge. Compared with other models (ResNet, DenseNet, or EfficientNet), our proposed method (Hybrid-ResUNet) had better efficiency and accuracy (Dice score: 0.93). Therefore, our federated learning framework effectively aids decision support systems across different hospital levels, protects patient privacy, and enables large-scale clinical trials. In addition, we propose the concept of human--robot collaboration to extract and transfer explainable features to deep learning models. The artificial intelligence technology described in this research could provide further insights into liver cancer and improve the quality of medical care. The automatic computation in image interpretation can also help reduce the workload of radiologists and facilitate liver cancer diagnosis.Funding: This work was supported in parts by the National Science and Technology Council (NSTC), Taiwan, under grants NSTC 111-2221-E-001-020 and 111-2321-B-075-004.Declaration of Interest: The Authors declare no Competing Financial or Non-Financial Interests.Ethical Approval: The protocol and the request for the waiver of informed consent for retrospective data collection and existing biosamples (REC No. 202109100RINC) have been approved by the 148th meeting of Research Ethics Committee C of the National Taiwan University Hospital. According to Good Clinical Practice guidelines and government laws and regulations, the following experiments are conducted under the committee. However, written informed consent is not required for participants in the study. Because the data set used for liver tumor segmentation in this paper is a public data set without patient identification, the LiTS (Liver Tumor Segmentation) data set is why consent is waived.
科研通智能强力驱动
Strongly Powered by AbleSci AI