Artificial Intelligence For Radiogenomics-based Tumor Survival Analysis

Purpose: Tumor is a major disease endangering human health.Survival prediction is of great significance for the personalized diagnosis and treatment of cancer patients.The rise of artificial intelligence technology provides new opportunities for survival predictive modeling.The intelligent prognostic model based on radiogenomics is the current frontier research topic.This topic combines radiogenomics and machine learning/deep learning methods to establish survival prediction models for myometrial invasive bladder cancer,medulloblastoma,and glioma,and verify their incremental value compared with existing clinical prognosis models.Materials and methods: 1.210 patients with muscle invasive bladder cancer were included,including 105 cases in training set and 105 cases in verification set.Using machine learning method,the radiomics signature based on Diffusion-Weighted Imaging,DWI)is established,and the image-clinical prognosis nomogram is generated.The incremental value of radiomics signature is evaluated in the aspects of model calibration,identification,reclassification and clinical usefulness.2.166 patients with medulloblastoma were included,including 83 cases in training set and 83 cases in verification set.RNA-seq sequencing data were collected from 17 patients.Machine learning method is used to establish the radiomics signature based on multi-parameter magnetic resonance imaging(MRI)sequence.Using radiogenomics analysis method,the key molecular pathways related to radiomics features were identified,and the prognostic value of pathway genes was verified in public data sets.3.A total of 1556 patients with glioma were enrolled in multicenter,which were divided into training set(935 cases),verification set(156 cases),the first external test set(194 cases),the second external test set(150 cases)and TCIA public data test set(121 cases).A DeepRisk convolutional neural network based on attention mechanism is constructed,which takes the undivided original image as input and outputs the survival risk of patients.Based on the DeepRisk score,the image-clinical prognosis nomogram was established to evaluate the incremental prognostic value of the DeepRisk score,and the performance comparison between the DeepRisk model and the traditional deep learning model was carried out.Results: 1.For muscular invasive bladder cancer,the radiomics signature is significantly correlated with progression-free survival(hazard ratio [HR] = 1.85,95%confidence interval [CI]: 0.96,3.56,p = 0.01),which is a prognostic factor independent of traditional clinical features(P < 0.001).Compared with clinical prognosis nomogram and radiomics signature,imaging-clinical prognosis nomogram has better performance in model calibration,identification,reclassification and clinical usefulness.2.For medulloblastoma,the radiomics signature is significantly correlated with the overall survival(HR = 4.36,CI = 1.66,11.43,P = 0.01)and the progression-free survival(HR= 2.46,CI = 1.06,5.72,P = 0.02),which is a prognostic factor independent of the traditional clinical features(P < 0.001).Compared with traditional clinical prognosis nomogram and radiomics signature,imaging-clinical pathology nomogram can better predict the overall survival(C-Index = 0.762)and progression-free survival(C-Index =0.697),and has better calibration and classification accuracy(NRI: OS: 0.298,P = 0.022;PFS: 0.252,P = 0.026).According to the analysis of radiogenomics,there are nine pathways related to radiomics features,including WNT signaling pathway,P53 signaling pathway,PI3K/AKT signaling pathway and so on,which are related to tumor prognosis.TCGA analysis showed that the average expression of pathway genes had significant prognostic value(HR = 0.63,CI = 0.38,1.06,P = 0.02).3.There is a significant correlation between DeepRisk model and overall survival(HR =13.48,CI =9.52,37.91,P < 0.001),which is a prognostic factor independent of traditional clinical features(P < 0.001).Compared with traditional Res Net model,the prediction accuracy of DeepRisk model is comparable,which shows that DeepRisk can realize accurate survival prediction by using undivided original MRI.After the DeepRisk model is incorporated into the clinical nomogram,the deep learning nomogram has better prognosis performance in calibration,identification,reclassification and clinical usefulness.Conclusions: 1.DWI-based radiomics signature is an independent prognostic factor for progression-free survival of patients with muscle invasive bladder cancer.Combined with radiomics signature,clinical stage and other clinical molecular pathological factors,it can better predict the progression-free survival of individuals,which confirms the incremental prognostic value of radiomics signature.2.Radiomics signature based on multi-parameter MRI sequences are independent prognostic factors for the overall survival and progression-free survival of patients with medulloblastoma,and are significantly related to key pathways.Combined with radiomics signature and other clinical molecular pathological factors,it can better predict the overall survival and progression-free survival of individuals,which confirms the incremental prognostic value of the radiomics signature.3.DeepRisk can learn and focus on the lesion area in MRI images of glioma patients.DeepRisk score is an independent prognostic factor in predicting the survival of glioma,and has incremental prognostic value compared with traditional clinical and molecular pathological risk factors.