Application Of Mr Radiomics In Predicting Extramural Vascular Invasion In Rectal Cancer

Objective:To investigate the application value of MRI-based imaging model in preoperative prediction of extramural vascular invasion of rectal cancer.Methods:A total of 202 patients who underwent pelvic MRI examination in our hospital from January 2019 to October 2022 and were pathologically confirmed as rectal cancer after surgery were selected,of which 63 were positive and 139 were negative.The patients were divided into the training set（141 cases）and the verification set（6 1 cases）in a ratio of 7:3.Clinical and imaging data of patients were collected and statistically significant（p≤0.05）data were included to establish clinical models.Receiver operating characteristic curve（ROC）was plotted and area under the curve（AUC）was calculated.In the radiomics analysis,Intelli Space Discovery（ISD）software was used to conduct semi-automatic 3D delineation of the primary lesion and the area of interest（ROI）of peritumoral vessels and lymph nodes along the lesion contour,and the 3 D volume of interest（VOI）of the lesion was generated automatically by the computer.The three-dimensional volume of interest delineated on the HR-T2WI sequence of the primary tumor lesion was referred to as T2 tumor lesion（VOI₁）for short.In the last phase of DCE-T1WI sequence enhancement,the diameter range of the outer edge of the intestinal wall of the focus was delineated to be about 4mm,and the three-dimensional volume of interest formed by the thickened and tortuous vessels outside the wall was referred to as DCE-T1 extrawall vessels（VOI₂）for short.The three-dimensional volume of interest delineated on the DWI sequence of lymph node formation around the rectum is referred to as DWI lymph node（VOI₃）for short.Four radiomic models and one combined clinical-radiomic model were established:T2 tumor lesion（Model 1）,DCE-T1 extramural vessel（model 2）,and DWI lymph node（model 3）,from the above three models,two models with better prediction efficiency were selected to form the joint image omics model（Model 4）;Then the model with the best predictive efficacy was selected from the four models to calculate the Radscore,and a combined clinics-imaging model（Model 5）was formed by combining the clinical data with the Radscore of the optimal model,and a nomogram was made.In the process,the Pyradiomics tool of Philips was used to extract the all radiomicss features of the outlined VOI.After dimensionless processing of the original image data,independent sample t-test was applied to the data conforming to normal distribution,and independent sample Mann-Whitney U test was applied to the non-normal distribution for preliminary screening.Then,the minimum absolute compression and selection algorithm（LASSO）in the machine learning model is applied to further reduce the dimension screening and select more meaningful features.Support vector machine SVM classifier was used to divide the sample data into training set and validation set in a ratio of 7:3.The ten-fold cross-validation method was used to draw ROC curve and calculate the value of AUC,sensitivity,specificity and accuracy to quantify the prediction efficiency.Finally,the model with the most predictive value was compared among the five models and a column graph was established.Calibration Curve and Decision Curve were used to evaluate the predictive performance of calibration curve.Calibration curve and decision curve were used to verify the clinical application value of calibration curve.Results:1.In clinical imaging data:mr T stage,mr N stage and mr EMVI of patients had statistical significance（p≤0.05）,and CEA in training set had statistical significance（p≤0.05）.2.DCE-T1 extramural vessel（Model 2）had the best predictive efficacy in the imaging omics model,with AUC value of 0.94 in the training set and 0.94 in the test set.DWI lymph nodes（Model 3）had the second best predictive efficacy,with AUC value of 0.93 in the training set and 0.91 in the test set.The combined clinical information-imaging feature model（Model 5）further improved the predictive efficacy,with the AUC value of 0.980 in the training set and 0.983 in the test set.3.Extramural vascular invasion probability of rectal cancer patients can be calculated based on the column diagram of Model 5.Conclusion:1.Model 5 has the best predictive ability of EMVI,and it shows good identification and calibration ability in both experimental and verification queues.2.DCA shows that the Model 5 to draw a nomogram has application value in clinical diagnosis and treatment.