Risk Prediction Of Radioimmune-associated Pneumonia In Non-small Cell Lung Cancer Based On A Multi-dimensional Deep Radiomics Model

Lung cancer is the leading cause of cancer-related death in the world,with 80-85%patients diagnosed with non-small cell lung cancer(NSCLC).The combination of immunotherapy and radiotherapy is playing an increasingly important role in the treatment of patients with NSCLC and has shown significant efficacy,but there is an increased risk of toxicity.Radioimmune-associated pneumonia was one of the most common adverse events associated with patient deaths.However,the occurrence and risk factors of radioimmune-associated pneumonia was controversial.Thus,it is necessary to develop an individualized model for assessing the occurrence risk of radioimmune-associated pneumonia.Deep learning provides us with a new research method for the extraction and fusion of high-throughput and multi-dimensional data,and is expected to provide a new tool for the individualized precision prediction of radioimmune-associated pneumonia and guiding clinical decision making.Part 1:Risk prediction of radioimmune-associated pneumonia in patients with NSCLC using a CT-based deep learning modelPurpose:To explore the application value of CT-based deep learning model in predicting the risk of radioimmune-associated pneumonia(grade≥2)in NSCLC patients.Methods:This study retrospectively collected 243 patients with NSCLC who received thoracic radiotherapy and immune checkpoint inhibitors(ICIs)from January 2015 to June 2021(62 symptomatic pneumonitis).Upsampling were performed in positive cases due to the uneven distribution of positive and negative patients.Five-fold cross-validation was performed for training and testing.The deep graph integrative model(DG)was composed of two pre-trained 3D UNet encoders to extract deep features from tumor and lung volumes of pre-treatment CT images,respectively,and a graph attention layer(GAT)for integration and classification.Results:Our new DG achieved area under ROC curve(AUC),sensitivity(sen),and specificity(spe)of 0.819,0.740,and 0.762,which outperformed conventional CT radiomics model(AUC 0.751,sen 0.713,spe 0.674),3D UNet based deep radiomics model(AUC 0.787,sen 0.735,spe 0.685),and our model without GAT(AUC 0.794,sen 0.735,spe 0.721).The improvement was statistically significant(p<0.001).Conclusions:Our CT-based DG model improved the prediction of the risk of symptomatic pneumonitis compared with traditional radiomics,which can be used as a new and effective tool for the screening of high-risk groups and the accurate assessment of the risk of pneumonitis.Part 2:Risk prediction of radioimmune-associated pneumonia in NSCLC patients based on an integrative multi-dimensional deep learning modelPurpose:To construct and validate a multi-dimensional fusion model based on deep neural network by integrating multimode and multi-source data for the prediction of radioimmune-associated pneumonia(grade≥2)in NSCLC patients,in order to further improve the model prediction performance.Methods:This study retrospectively collected 213 patients with NSCLC who received thoracic radiotherapy and ICIs from January 2015 to June 2021,of whom 59 patients developed symptomatic pneumonitis.Patient clinical data,blood inflammatory biomarkers,dosimetric parameters and pre-treatment CT images were collected.Upsampling were performed in positive cases due to the uneven distribution of positive and negative patients.Five-fold cross-validation were performed.The adaptive fusion and prediction model consists of a)3D CNN encoders for deep feature extraction from tumor and lung volumes,b)multilayer perceptron(MLP)for deep signature extraction from categorical and numerical clinical parameters,and c)graph attention layers(GAT)to learn adaptive wegihts of different types of features for embedding,and d)softmax for final classification.Results:Our multi-dimensional prediction model achieved AUC,sensitivity(sen),and specificity(spe)of 0.859,0.799,and 0.714,which improved the performance of other comparing methods including our model using CT alone(AUC 0.816,sen 0.731,and spe 0.714),our model without GAT(AUC 0.841,sen 0.773,and spe 0.688),and conventional radiomics model based on logistic regression(AUC 0.794,sen 0.740,spe 0.708).The improvement was statistically significant(p<0.001).Conclusions:This study constructed a multi-dimensional fusion model based on deep neural network utilizing multimode and multi-source,which maximizes the model performance and achieves accurate prediction of radioimmune-associated pneumonia,providing a potentially useful and effective tool to assist the safety management and individualized treatment of NSCLC patients.