Construction Of A Prognostic Risk Score Model With Multimodal Features Of ¹⁸F-FDG PET/CT Images Of Non-small Cell Lung Cancer By Features Fusing

Objective:Through a small multicenter dataset,we extracted radiomics features and deep learning features from PET/CT images of different modalities to construct machine learning models for feature fusion,obtained their prediction scores as input features,combined with clinical information,and constructed a Cox proportional risk regression model as the prediction model for the risk of death scores of patients with non-small cell lung cancer,to provide support for clinical decisions such as early detection of patients and improvement of patient prognosis.Methods:In this study,we used the ¹⁸F-FDG PET/CT public dataset（98 cases）as the training set and our data as the validation set（41 cases）.The 3D lesions were firstly outlined to extract the radiomics features,and then the Res Net50 and VGG16 networks were constructed by transfer learning to extract the deep learning features at the largest lesions of 2D PET and CT images.After feature selection by Lasso method,RSF,XGBoost,GBM,Coxboost,SSVM,and Deepsurv survival analysis machine learning models were constructed to fuse high-dimensional features,respectively.The best performance models were selected to output predictive values combined with clinical data（gender,age,smoking history,pathological staging and TNM stage）to construct risk score prediction models by univariate and multifactor Cox analysis.Results:139 patients were included in this study.Outcome events occurred in 43 of them（28 in the training set and 15 in the test set.）ROC analysis showed that the GBM model constructed using CT and PET radiomics features and deep learning features performed best,and it was selected to fuse various types of features and output predictive values,and after univariate and multifactor Cox analysis by combining clinical features to select T-stage（HR=1.680）,CT-radiomic-GBM predictive score（HR=2.759）,PET-radiomic-GBM predictive score（HR=2.539）,CT-Res Net50-GBM predictive score（HR=7.778）,and PET-VGG16-GBM predictive score（HR=3.317）were selected as predictors to construct the Cox proportional risk model.The model had good performance in the test set（AUC=0.738±0.072）and could bring more benefits for clinical decision making compared with the traditional TNM staging model.Conclusion:By fusing high-dimensional features through machine learning models,deep learning features can effectively complement radiomics features to combine anatomical and metabolic information from PET/CT images with clinical data to predict the prognosis of patients with non-small cell lung cancer.The risk score model constructed in this study showed strong predictive ability compared to conventional TNM staging.