信息工程与人工智能学院知识库

Non-small cell lung cancer (NSCLC) is the leading cause of lung cancer-related death. Survival risk stratification of NSCLC patients can help doctors develop individualized treatment plans, plan follow-up plans and prolong patients' survival. Artificial intelligence can identify key information from a large amount of medical information to help patients' prognosis. Tumor, node, and metastasis（TNM） staging is the main means of survival risk stratification, and also an important indicator for doctors to judge patients' survival risk. However, TNM staging requires pathological detection, which may bring the risk of infection to patients. To solve this problem, in this paper, positron emission tomography/computed tomography (PET/CT) images were used to obtain TNM staging of NSCLC patients and stratify survival risk of NSCLC patients, so as to reduce the pain of patients' pathological testing and help doctors make effective decisions. The main research is as follows:

(1) In order to accurately stratify the survival risk of patients with NSCLC, a deep learning model was established for 467 patients with NSCLC through PET/CT images before treatment, and the patients were stratified according to the survival prediction results. First, the original PET/CT images were processed by preprocessing technology to obtain the full-body maximum density projection (MIP) images of the patient in front and side. Secondly, based on MIP image, radiomics data and clinical information, a multi-modal information fusion module is designed in the deep learning model. The deep learning model was then trained to obtain survival risk stratification. Finally, the area under receiver operation characteristic curve (AUC) and accuracy were used to evaluate the model performance. In predicting patients' survival risk, the model had an AUC of 0.84 and an accuracy of 0.78, higher than that of the random forest model with clinical information input (accuracy: 0.78, AUC: 0.78) and the traditional radiomics model (accuracy: 0.71, AUC: 0.78). The model presented in this paper can effectively stratify the survival risk of NSCLC patients and provide early prognostic information for patients.

(2) In order to obtain TNM staging and survival risk stratification of NSCLC patients, this paper proposes an end-to-end multi-task deep learning model. In the deep learning model framework of this paper, the level of TNM staging and survival risk stratification can be predicted. The multi-task deep learning model takes TNM staging as the auxiliary task and survival risk stratification of NSCLC patients as the main task. The two tasks promote each other in the training process. First, the deep learning features of the front and side MIP images were extracted using Swin Transformer and fused with the radiomics features. Post-fusion features were then used to simultaneously predict TNM stage and survival risk stratification in patients with NSCLC. The experimental results show that the multi-task deep learning model proposed in this paper is superior to the traditional stochastic forest model based on radiomics and the single-task deep learning model.