Application Of Radiogenomics In The Prognosis Of Colorectal Cancer

Radiogenomics can explain medical image features at the biological level.As an input to predict patient prognosis,imaging features can be extracted using traditional feature extraction software such as 3D Slicer and Pyradiomics or deep learning algorithms.There were few related studies on combining the obtained two types of features and predicting the prognostic risk level of patients.In this study,the signaling pathways related to prognosis were used to annotate the image phenotype at the biological function level;through the image data set dynamically generated by the transcriptome and image feature labels,a deep learning model based on ResNet18 was constructed to predict the prognosis risk level of colorectal cancer.This study focuses on how to establish a relationship between imaging feature phenotypes and biological functions.In addition,the feasibility of converting the transcriptome one-dimensional data into two-dimensional image data as the input of the colorectal cancer prognosis risk grade prediction model for high and low prognosis classification was explored.The data sources are the colorectal cancer data of Dazhou Central Hospital in Sichuan Province and the colorectal cancer data of TCGA(The Cancer Genome Atlas)public database.First,we used the Radiomics data set containing 215 people,and used single factor Cox and LASSO-Cox(Least Absolute Shrinkage and Selection Operator-Cox)analysis to obtain the imaging features related to prognosis.Next,the WGCNA(Weighted Gene Co-expression Network Analysis)was performed using the Radiogenomics dataset of 83 people,and the resulting gene modules were combined with the TCGA dataset of 467 people to search for robust modules based on Zsummary.GSVA(Gene Set Variation Analysis)was used to perform an enrichment analysis of the genes in the gene module that were significantly associated with the prognostic imaging features.The biological annotation of the prognosis-related imaging phenotypes obtained above was realized by using the pathway information significantly correlated with the prognosis imaging features.Finally,the genes in the signaling pathways that were significantly related to the prognosis image features will be extracted,and the one-dimensional gene expression was converted into two-dimensional artificial image objects as the input of the ResNet18 deep learning model for the binary classification task of prognostic risk level,then use the Grad-CAM algorithm to identify the decision-making basis of the prognosis risk level prediction model in the form of highlighting in the image.26 prognosis-related image features were obtained from the Radiomics dataset through the image feature selector;13 highly robust modules(Z >10)were obtained from the Radiogenomics dataset through Zsummary,4 of which were significantly correlated with prognosis-related image features(|R| > 0.2 & P < 0.05).After GSVA pathway enrichment for the genes in the 4 modules,124 pathways were found to be significantly correlated with23(23/26)image features(|R| > 0.4 & P < 0.01)and most pathways(73/124)were found to be significantly correlated with immune related.In the task of predicting the prognostic risk level of the ResNet18 model,the Accuracy and AUC(Area Under Curve)in the training set were 0.84 and 0.94,respectively,and the Accuracy and AUC in the test set were0.81 and 0.83,respectively.In this study,23 imaging features were annotated with biological functions and found to have immune-related functions.In addition,by converting transcriptome data into artificial image objects,a model for predicting prognostic risk levels was constructed using ResNet18.Realize the function of predicting high and low risk of patient prognosis based on imaging and transcriptome data.The attempt to convert one-dimensional transcriptomes into two-dimensional images provides ideas for subsequent sequencing data mining.