Research On Radiomics Technology With Application In Clinical Tumor Diagnosis And Treatment

BACKGROUND: In clinical oncology research based on medical image analysis,Radiomics technology has developed rapidly since its appearance and has become a relatively mature technology that is wide.It has been applied to the diagnosis and treatment of various clinical tumors researches.Radiomics technology can quantify the tumor information in medical images with high throughput and performs feature analysis,modeling,and verification in a physically explicable way,to build diagnostic and therapeutic prediction tools that are easy to use clinically.At present,Radiomics has successfully become one of the most advanced research topics for radiation oncologists,radiologists,and doctors in other departments at home and abroad,but it is still difficult for them to complete the whole experiment independently.As a novel signal processing technology(Radiomics),it is necessary for us to systematically summarize the research framework and the key technical links for the convenience of researchers in the field of signal processing and medical science,from the point of view of signal and information processing.At the same time,there are still many clinical cancer problems that have not been studied,and it is necessary for us to carry out individualized prediction research based on domestic cancer patients,in order to promote the development of intelligent cancer diagnosis and treatment in China.Therefore,the purpose of our study mainly includes:(1)A systematic overview of Radiomics technology,with summarizing the research framework and key technical links;(2)Carry out the application research of Radiomics technology in different clinical cancer diagnoses and treatment,to explore the potential value of Radiomics in clinical application.METHODS:(1)Prediction of resistance to neoadjuvant chemoradiotherapy(NCRT)in locally advanced rectal cancer(LACR)based on ADC Radiomics.The pre-treatment ADC images of 89 patients with LARC were retrospectively analyzed,and 133 Radiomics features were extracted based on the region-of-interest(ROI).Then,feature importance was evaluated by the Random Forest algorithm combined with 9 clinical features.The first 10 features were selected to construct the Random Forest prediction model in the training dataset and tested in the test dataset.(2)Prediction of response to Gamma Knife Radiosurgery(GKRS)in lung cancer brain metastases(LCBM)based on multimodality MRI Radiomics.The data of 213 lesions of 137 patients were retrospectively analyzed.There were seven pre-treatment MRI sequences of each lesion,and the core area and the peritumoral edema area were delineated,respectively.And the omics features were extracted respectively.Then,the Radiomics prediction performance of models was compared between different regions and different MRI sequences.Finally,the random forest radiomics score and nomogram were developed in the training cohort and tested in an independent validation cohort divided according to the time sequence.(3)Prediction of PD-L1 expression levels in non-small cell lung cancer(NSCLC)based on CT Radiomics.Pre-treatment CT images of 125 patients from two centers were retrospectively analyzed,and 1287 quantitative features were extracted based on the delineated tumor region of each data.Ridge Regression-based feature recursive elimination(RFE)algorithm was used to implement feature selection.Logistic Regression was used to develop and compare the predictive potential between the Radiomics Score,clinical model,and Radiomics-clinical nomogram in the training cohort,and three models all were tested in an independent validation cohort.(4)Preoperative classification of glioblastoma(GBM)and brain metastasis(BM)based on T1/T2/T1 CE Radiomics.Preoperative T1/T2-weighted images(T1WI/T2WI)and contrastenhanced T1WI(T1CE)of 140 patients with GBM and 128 patients with BM were retrospectively analyzed.Hand-crafted Radiomics(HCR)and deep learning-based Radiomics(DLR)were used to analyze the tumor regions.Different machine learning models were used to determine the best modeling method,and the prediction performance of different combination modes between two kinds of radiomics features and different modality MRI were compared.RESULTS:(1)In the first study,a total of 10 valuable quantitative characteristics were mined,and the interclass correlation coefficients(ICCs)range from 0.716 to 0.948,which shows good reproducibility.In the training dataset,the area under the curve(AUC)of the receiver operating characteristic(ROC)curve for the identification of resistance to NCRT was 0.84.In the test dataset,the AUC was 0.83,the highest accuracy was 91.3%,the sensitivity was 88.9%,and the specificity was 92.8%.(2)In the second study,we found that sex,histological subtype,epidermal growth factor receptor mutation,and targeted drug treatment were significantly associated with posttreatment response(P = 0.009-0.034).In the comparison of different MRI sequences,T2WI(AUC = 0.725),T2-FLAIR(AUC = 0.704)and CBV(AUC = 0.714)showed the better prediction performance.In addition,adding Radiomics features of peritumoral edema(RFPE)to Radiomics features of tumor core(RFTC)showed improved prediction performance than RFTC alone(AUC=0.848 versus AUC=0.750;P<0.001).Finally,the Radiomics nomogram based on Radiomics Score and clinical factors reached the highest prediction performance(training cohort,AUC=0.930;validation cohort,AUC=0.852).(3)In the third study,there was no significant difference in PD-L1 expression levels determined by clinical characteristics(P = 0.109–0.955).Upon selecting 9 radiomics features,we found that the Logistic Regression-based prediction model performed the best(AUC = 0.96,P <0.001).In the external validation cohort,our Radiomics Score showed an AUC of 0.85,which outperformed both the clinical model(AUC = 0.38,P < 0.001)and the Radiomics-clinical nomogram model(AUC = 0.61,P < 0.001).(4)In the fourth study,Random Forest was the optimal modeling method for HCR and DLR.HCR models already showed good results for distinguishing between the two types of brain tumors in the test dataset(T1WI,AUC?=?0.86;T2WI,AUC?=?0.76;T1CE,AUC?=?0.93).By adding DLR features,all AUCs showed significant improvement(T1WI,AUC?=?0.87;T2WI,AUC?=?0.80;T1CE,AUC?=?0.97;P< 0.05).The T1CE-based HCR + DLR model showed the best classification performance(AUC?=?0.99 in the training dataset and AUC?=?0.97 in the test dataset),surpassing the other MRI modalities(P< 0.05).CONCLUSIONS: Based on the research framework and methods of Radiomics,four clinical studies on the diagnosis and treatment of cancer were carried out.Preliminary results showed the application potential of Radiomics in the prediction of resistance to NCRT in LACR,the prediction of response to GKRS in LCBM,the prediction of PD-L1 expression levels in NSCLC,and the classification of GBM and solitary BM.All four studies have confirmed the application value of Radiomics in the clinical diagnosis and treatment of cancer.