This study was approved by the Ethics Review Committee of the First Affiliated Hospital of Henan University of Science and Technology, and all procedures were performed in accordance with the principles of the Helsinki Declaration. We retrospectively enrolled 412 patients with parotid tumours undergoing MRI examination at the First Affiliated Hospital of Henan University of Science and Technology between 2013 and 2019. The following inclusion criteria were used: 1) patients received no treatment before the examination; 2) the T1WI and T2WI sequences of the MRI scans were complete and available; 3) the images were clear and without artefacts; 4) a definite pathological diagnosis by surgery and pathology was provided for the patients. Finally, data from 112 PA patients and 140 PPA patients were collected in this study.
The clinical features of the 252 subjects are listed in Table 1. Among the PA patients, the average age was 55.57±1.29 years (range: 23-77 years) and the gender ratio (M:F) was 1.38:1. Among the PPA patients, the average age was 47.81 ±1.473 years (range: 15-81 years) and the gender ratio (M:F) was 0.67:1. All 252 subjects were randomly allocated to the training cohorts and validation cohorts at a ratio of 7:3, according to previous published reports [19,26]. Therefore, 176 cases were assigned to the training cohort (PA/PPA = 78/98) and the other 76 patients were assigned to the validation cohort (PA/PPA = 34/42). The flow chart of the procedure is given in Fig. 1.
All subjects underwent routine 1.5 Tesla MRI scanning (GE Signa HDX 1.5 T; GE Healthcare, Milwaukee, WI) with a head-neck coil. The scanning sequence was acquired including the fast spin echo T1WI and the fast spin echo T2WI with fat saturation. The parameters of T1WI were: TR of 700.0 ms, TE of 8.9 ms, matrix size of 320×192 mm, FSE of 24 cm×24 cm, slice thickness of 5 mm, spacing of 1 mm. The parameters of the T2WI sequence were: TR of 3900.0 ms, TE of 100.0 ms, matrix size of 320×256, FSE of 24 cm×24 cm, slice thickness of 5 mm, slice spacing of 1 mm in the axial images; TR of 3300.0 ms, TE of 100.0 ms, matrix size of 320×224, FSE 24 cm× 24 cm, slice thickness of 5 mm, slice spacing of 1 mm in coronal images.
MRI imaging data came from our organization's image archiving and communication system (PACS). Two board-certified senior radiologists (readers 1 and 2, with 8 and 13 years of clinical experience in head and neck diagnosis, respectively) independently interpreted the MRI images (including the T1WI and T2WI sequence scans) in the PACS of the Radiology Department (Fig. 2 a-b). The two radiologists manually delineated the ROI (region of interest) by using MATLAB (2014b, MathWorks, Natick, MA, USA) and an open source program software, Imaging Biomarker Explorer (IBEX, http://bit.ly/IBEX_MD Anderson). The extracted features included the intensity histogram, grey co-occurrence matrix (GLCM), grey run length matrix (GLRLM) and shape (Additional file 1). Reader 1 extracted features twice with the same procedure, which were used to measure the intra-observer consistency. At the same time, reader 2 extracted features independently, and the feature data collected by reader 2 were compared with those obtained by reader 1 to evaluate inter-observer consistency. The intraclass correlation coefficient (ICC) was used to calculate the consistency, and the features with robust consistency (ICC> 0.75 for both in the intra-observer and inter-observer rates) were retained for subsequent selection.
Radiomics feature selection
After z-score normalization, the extracted features (ICC>0.75) of the T1WI and T2WI sequences were examined by an independent sample t-test (continuity variable) or a Mann-Whitney U test (classified variable). Here, the selected features of the T1WI and T2WI sequences (p<0.05) were combined as the T1-2WI features. All of the retained T1WI, T2WI and T1-2WI features were processed by dimensionality reduction using the LASSO method to improve the accuracy and degree of modelling fit . Data within 1-standard error of the minimum criterion measure were used in this study. Then, the correlation coefficients of the radiological features were assessed by Spearman analysis, and the radiological features with high linear correlations (correlation coefficients of 0.90-1.00) were excluded.
Construction of the radiomics models
After the dimensionality reduction procedure, the important and independent T1WI, T2WI and T1-2WI features were separately used to construct radiomics models by two machine learning methods (MLR and SVM). The discriminatory performance of the models was quantified and evaluated in the training and validation cohorts according to the AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F-1 score. The calibration of the radiomics model was calculated by the Hosmer-Lemeshow test. The independent clinical feature model was established with the clinical features by MLR. Then, the diagnostic efficacy was compared between the radiomics model and the clinical feature model for both the training cohort and the validation cohort.
R (version 3.4.1, https://www.r-project.org/) was used for the statistical analysis. The normality of the distribution and the homogeneity of the variance were evaluated by the Shapiro-Wilk test and Bartlett’s test, respectively. Continuous variables were compared by independent t-tests or Wilcoxon rank sum test, while categorical variables were compared by chi-square or Fisher’s exact test. LASSO regression was carried out using the “glmnet” package with multivariate binary logistic regression. The correlation coefficient matrix was visualized by the “ggplot2” and “ggcorrplot” packages. SVM models and ROC curves were generated with the “e1071” and “pROC” packages, respectively. The AUCs were compared using the “DeLong” test in both the MLR and SVM models. A P-value<0.05 indicated a significant difference.