Cohort CT image features and knowledge-driven lung nodule risk assessment
and reduce the medical burden. However, a large number of pulmonary nodules were screened with various categories and complex morphology, and the sensitivity and specificity of existing screening methods for the diagnosis of malignant risk cannot meet the clinical needs, and over or under diagnosis treatments are common. The main reason is that it is difficult to predict the growth trend and risk of nodules. The computational model driven by the combination of regular follow-up CT image feature data and clinical knowledge provides a new idea to break through the above difficulties. Based on the cohort chest CT image data obtained during follow-up, this project intends to deeply mine the imaging information of pulmonary nodules based on AI technology, overcome the problem that the cohort CT images cannot be fully matched due to the patient's position, the degree of lung inflation and other reasons, and introduce clinical information and expert knowledge to study the risk assessment of pulmonary nodules. We plan to carry out the following innovative works: specific analysis of lung nodules and
classification ; Cohort CT image based lung nodule growth prediction; Construction of lung nodule risk assessment model by fusing image features, knowledge graph and clinical information. It is expected that the research results will contribute to the personalized and precise diagnosis and treatment of early screening of lung cancer, which has important research value and application prospect.
This project intends to build a lung nodule risk assessment model in lung cancer screening through theoretical or technological breakthroughs such as classification and segmentation of lung nodules and growth prediction, as well as innovations in multi-source data and knowledge fusion modeling technology. Specific research objectives include:
1) In view of the diversified forms and small proportion of malignant nodules in lung cancer screening, subclass annotation data sets were established to obtain specific characteristics of nodules on the basis of nodule segmentation, and the model was guided to achieve high-performance benign and malignant prediction and classification results of lung nodules under the small-sample learning paradigm.
2) Construct CT data set of lung cancer screening cohort, study lung nodule pairing and growth prediction algorithm in CT images of cohort, and solve image differences caused by patient position, lung inflation degree and scanning and acquisition equipment at different time points.
4) The medical knowledge map of lung cancer diagnosis and treatment was constructed, and the lung nodule risk assessment model was constructed by combining the knowledge characteristics of lung nodule patients with the quantitative feature database of multi-source data, so as to explore the potential correlation between the image characteristics of patients with lung nodule and clinical information.
Bin Zheng，Fujian Medical University
Shaohua Zheng，Fuzhou University
Liqin Huang，Fuzhou University
Taidui Zeng，Fujian Medical University
Lin Pan，Fuzhou University
Wei Li，Fuzhou University