| Objective: Construction of automatic semantic segmentation model for primary lesion of renal clear cell carcinoma based on deep learning.Constructing a machine learning model for predicting tumor grade of renal clear cell carcinoma based on image genomics features.Analysis of genes associated with renal clear cell carcinoma opens up new possibilities for individualized precision medicine.Methods: This is a retrospective study based on public database.The Cancer Imaging Archive(TCIA)and The Cancer Genome Atlas(TCGA)renal clear cell carcinoma data set(TCGA-KIRC)were downloaded and collected.According to the clinical information of data set and the World Health Organization & International Society of Urology(WHO/ISUP)classification in 2016,the tumor clinic was divided into low-risk group and high-risk group,including 88 cases in high-risk group and109 cases in low-risk group.The U Net deep learning network model for automatic segmentation of tumor primary focus is constructed by Tensorflow and Keras platforms respectively.And use PyRadiomic library in Python(3.8.5)software to extract image omics features from the results of automatic segmentation model.Wilson’s rank sum test was used to screen out the image features and expression genes which were significantly different from different tumor stages(P<0.05),and Fisher was used to accurately test the screened mutant genes.Aiming at the features with obvious differences,the dimension of different omics features is reduced by using the Least Absolute Shrinkage and Selection Operator(LASSO)respectively.According to the ratio of 8:2,the samples were randomly divided into training set and test set.Then,according to different omics characteristics,the machine learning model was constructed and the receiver operating characteristic curve(ROC)and the Area Under Curve(AUC)were calculated.Results: The score of the deep learning model with cortical segmentation and the score of the deep learning model with medullary segmentation were 98.46% and99.03%,respectively,in the U Net neural network constructed during the experiment.According to the automatic segmentation results,107 radiomics features were extracted.After statistical analysis,24 image features,2125 expressed genes and 37 mutated genes with significant differences in different tumor grades were obtained.The area under the curve(AUC)of the classifier constructed by the different omics characteristics was as follows: AUC of the predictive model constructed by radiomics was 71.5%(95% CI: 55.1%-87.8%),AUC of the predictive model constructed by gene expression data was 85.6%(95% CI: 73.2%-98%),and AUC of the predictive model constructed by gene mutation data was 65.2%(95% CI: 47.8%-82.5%).The AUC of the composite predictive model constructed with the integration of radio-genomic features was 92.9%(95% CI: 0.841-1.18).In addition,functional enrichment analysis showed that some key genes of renal clear cell carcinoma were enriched in multiple functional pathways,such as correlation between WNT4 and T cell differentiation,regulation of cell matrix adhesion,organelle fusion,etc.We observed that among the significantly different mutation genes.Conclusion:The deep learning model developed and constructed in the research process can accurately segment the primary lesion of renal clear cell carcinoma.The composite classifier constructed by machine learning was more efficient than the single omic classifier in predicting the different risk subgroups of renal clear cell carcinoma.The analysis of the results of this study has revealed that key genes such as WNT4,BAP1 and SETD2 are involved in the occurrence and development of tumors and are expected to serve as potential biological markers of renal clear cell carcinoma. |
PDF Full Text Request