Patients
This retrospective study was approved by the ethics committee of our hospital. The required informed patient consent was waived. We collected patients who underwent surgical treatment from May 2019 to December 2019 for model construction (train cohort), who underwent surgical treatment from January 2020 to August 2020 for validation (validation cohort), and who underwent surgical treatment from September 2020 to July 2021 for test (test cohort). All patients underwent pelvic DCE MRI before surgery.
The inclusion criteria were as follows: (1) primary tumor was histopathologically proven to be rectal cancer; (2) postoperative pathological diagnosis of EMVI was confirmed; (3) preoperative pelvic DCE MRI was performed; and (4) patients had never been treated with neoadjuvant chemoradiotherapy. The exclusion criteria were as follows: (1) primary tumor was not clearly visible on MR images because of artifacts, (2) primary tumor displayed incompletely on DCE images, and (3) other malignancies were present. The pathological EMVI was diagnosed by the consecutive sections of the entire en-bloc specimen when tumor cells were found contacting tightly with endothelial cells in vasculature, including blood and lymph vessels, without specifying the intra- or extra-mural invasion. And then the patients were divided into 2 groups by the pathological EMVI: EMVI-positive (n = 17, 3, 5 in train, validation and test cohorts, respectively) and EMVI-negative (n = 27, 4, 13 in train, validation and test cohorts, respectively).
We also gathered the following patients’ clinical characteristics: sex, age, body mass index (BMI), carcinoembryonic antigen (CEA), and carbohydrate antigen 19-9 (CA19-9). BMI, an indicator of body fat, is calculated using the individual's weight and height according to the following formula: \({\text{BMI}}\left( {{\text{kg/m}}^{2} } \right) = {\text{Weight}}\left( {{\text{kg}}} \right)/\left( {{\text{Height}}\left( {\text{m}} \right)} \right)^{2}\).
MRI data acquisition
All patients underwent preoperative pelvic MRI scans in the supine position on a 1.5-Tesla MR scanner (Avanto, Siemens, German) with a phased-array body coil. The scan sequences included T1WI (transverse position), T2WI (transverse and sagittal position), high-resolution T2WI (axial position of abnormal intestinal segment), DWI (b-value of 0 and 800 s/mm2) and DCE-MRI. The DCE-MRI adopted 3D-volume interpolated examination (3D-VIBE) and the scan parameters were as follows: multiple flip angles of 5°, 10°, 15°, respectively; each angle was scanned one time, each acquisition time was 8 s, total time was 24 s; repetition time (TR) = 4.88 ms; echo time (TE) = 2.39 ms; Average = 1; field-of-view (FOV) = 350 × 280 mm2; matrix = 288 × 196; slice thickness = 4.0 mm; and bandwidth (BW) = 400 Hz/px. The DCE parameters were the same as the multi-angle parameter; the flip angle was 10°; multiphase dynamic enhancement scanned 30 phases; the imaging time was 4 min. We injected 0.1 mmol/kg gadolinium diamide (Omniscan; GE Healthcare, Little Chalfont, UK, http: //www. gehealthcare.com) in the center of the elbow using the high-pressure vein syringe and then injected 20 mL of physiological saline at the speed of 2 mL/s. The scan parameters of the other scan sequences were described in the Additional file 1: Appendix A1.
All MR images were retrieved from the picture archiving and communication system (eWorld, China).
Conventional MRI assessment
Two radiologists with 10 (W.L.) and 15 (J.F.) years of experience in abdominal radiology assessed rectal cancer based on MRI individually. The assessment included the distance between the rectal cancer and anal edge (distance), MRI-based T staging (cT), MRI-based regional lymph node metastasis assessment (cN), number of visible regional lymph nodes on MR images (LN), MRI-based EMVI assessment (MR-EMVI) and MRI-based CRM assessment (MR-CRM). Disagreements were resolved by consensus or consultation with a third radiologist (G.T.) with over 25 years of experience in abdominal radiology. All radiologists were blinded to the histological results.
Data processing and tumor segmentation
DCE MRI images were transferred to quantitative Omni kinetics software (OK, GE Healthcare, China). First, the arterial input function (AIF) was placed on the proximal abdominal aorta and the AIF curve was obtained. Then, the extended Tofts linear model was selected to obtain the pharmacokinetic parameters Ktrans, Kep, Ve, and Vp (Ktrans: volume transfer constant between the blood plasma and the extracellular extravascular space (EES), Kep: rate constant of contrast agent escapes from the EES into the plasma compartment, Ve: EES volume fraction, and Vp: plasma volume fraction). Three-dimensional tumor segmentation was performed by two radiologists D.W. and W.S. The region of interest (ROI), covering the whole primary tumor without the adjacent tissue, lumen, or intestinal content, was outlined on the original images first and then copied into the Ktrans, Kep, Ve and Vp maps.
The texture features of the above pharmacokinetic parameters were extracted by OK software, including histogram features (number = 12), grey-level co-occurrence matrix (GLCM) features (number = 13), Haralick features (number = 9), grey-level run-length matrix (RLM) features (number = 16), and formfactor features (number = 9). Fifty-nine features were computed using every pharmacokinetic parameter, and a total of 236 features were obtained.
To evaluate the feature reproducibility across different radiologists, 30 cases were randomly selected for a double-blinded comparison of manual segmentations by the two radiologists. The inter-observer agreement of the features was evaluated using the interclass correlation coefficient (ICC). An ICC of > 0.75 was considered as a mark of excellent reliability.
Texture features selection
To eliminate the differences in the value scales of extraction features, feature normalization was performed before feature selection. Each feature for all patients was standardized using Z-scores.
Analysis of variance (ANOVA) and Kruskal–Wallis tests were performed to select the texture features significantly correlated with EMVI. Then, the least absolute shrinkage and selection operator (LASSO) regression method by penalty parameter tuning λ were used to reduce the redundancy or selection bias of the features. Optimal λ was chosen based on the minimum criteria in a tenfold cross-validation. This method is widely used for high-dimensional features with small medical images.
After feature selection, texture score (T-score) was derived from the linear combination of the selected features that were weighted by their respective LASSO coefficients to reflect the probability of EMVI for each patient. The predictive capability was evaluated using the receiver operating characteristic (ROC) curve and area under the curve (AUC).
Assessment model construction
Univariate logistic regression was used for clinical information and conventional MRI features to select independent clinical predictors. Multivariable logistic regression analysis with the selected independent clinical risk factors and T-score was applied to develop a combined model for the EMVI assessment model. Then, we used the validation cohort to conduct a preliminary assessment of the model.
A backward stepwise multivariable logistic regression was performed using the Akaike information criterion (AIC) as the stop rule. To provide the clinician with a quantitative tool for predicting the individual probability of EMVI, a nomogram was plotted based on the EMVI assessment model. The differences between the ROC curves of MR-EMVI, T-score, and the assessment model were compared using the DeLong test.
Finally, the assessment model was tested by a test cohort.
Clinical practice
To estimate the incremental utility of the T-score and assessment nomogram model, the decision curve of the different models was plotted for the entire dataset. The decision curve analysis (DCA) informs the patient or doctor which of the several models are optimal using a threshold probability.
Statistical analysis
All statistical analyses were performed using R software (version 3.5.1; http://www.Rproject.org). Univariate analysis for clinical features was implemented using an independent-sample t test or the Mann–Whitney U test for continuous variables, and the Chi-squared test for categorical variables. All statistical significance levels were two-sided, with statistical significance set at 0.05. The LASSO logistic regression was conducted using the “glmnet” package in R software. Multivariate logistic regression analysis was performed using the “stats” package. Lastly, nomogram construction was done using the “rms” package.