Patient
This retrospective research was approved by our hospital ethics committee, and patient informed consent was obtained. This retrospective investigation was authorized by our hospital's ethics committee, and patient consent was obtained. The study's inclusion criteria comprised two factors: 1) confirmation of PCNSL or SBM through histopathological analysis and 2) the completion of preoperative MRI. Conversely, the exclusion criteria included 1) the absence of contrast-enhanced T1-weighted imaging (CE-T1WI) sequences, 2) prior receipt of antitumor therapy, such as surgery, radiotherapy, or chemotherapy before MRI scanning, and 3) the presence of significant artifacts in the images, as determined by radiologists with a decade of experience. Ultimately, the study included a total of 115 patients with pathologically confirmed PCNSL and 209 patients with pathologically confirmed SBM from our hospital between January 2011 and May 2021.
The patient selection process is shown in Fig. 1.
MRI Acquisition Protocol
In this study, all participants underwent examination utilizing 3.0T Siemens Trio Scanners located in the radiology department of the institution. Dadopentetate dimeglumine was administered intravenously at a dose based on the patient's body weight (0.1 mmol/kg) as a contrast agent. Axial CE-T1WI images were acquired using the MPRAGE sequence with the following imaging parameters: TR/TE/TI = 1900/2.26/900 ms, Flip angle = 9°, slice thickness = 1 mm, axial FOV = 25.6 × 25.6 cm2 and data matrix = 256 ×256. Multidirectional data for CE-T1WI images were obtained during the interval time of 90–250 s.
Image Segmentation
The open-source software 3D Slicer facilitates three-dimensional manual segmentation (available at https://download.slicer.org/). The manual segmentation of tumor tissues on axial CE-T1WI was conducted by a solitary, experienced radiologist with over seven years of expertise in neuroimaging. Adhering to the software protocol, the radiologist delineated the complete lesion for each image layer to obtain tumor segmentations. The regions of interest (ROIs) were meticulously plotted along the tumor tissue boundary in each slice, encompassing the necrosis and neoplastic blood vessels, while excluding the peritumoral edema. Each segmentation was validated by a senior radiologist with a decade of experience. The radiographic lesions of a classic nature are depicted in Fig. 2, while the workflow of the radiomics analysis is presented in Fig. 3.
Radiomic Feature Extraction
PyRadiomics open-source software was utilized to extract imaging features of the tumor after segmentation of the entire tumor. A standard normal distribution of the image distribution was obtained through standardized work on axial CE-T1WI images.
A comprehensive set of 1561 radiomics features, comprising both original features and secondary features derived through LoG, exponential, square, square root, logarithm, and wavelet transformations of the primary features, were extracted from the axial CE-T1WI images of the tumor. These extracted features were employed to quantify diverse tumor characteristics, encompassing size (e.g., volume), shape (e.g., circumference, diameter), grayscale cooccurrence matrices (e.g., energy, contrast, entropy), grayscale run-length matrices, grayscale dependency matrices, and other relevant parameters.
Radiomics Feature Selection and Development of Radiomics Models
The construction of a diagnostic model comprises two components, namely, radiomics feature selection and the deployment of a classification algorithm or classifier. The objective of feature selection is to reduce the dimensionality of the original features, enhance the model's generalizability, and fortify the assembly process. While imaging features offer insights into tumor characteristics from diverse perspectives, not all information is closely interrelated.
A two-step feature selection process was employed, beginning with preselection of candidate features using the least absolute shrinkage and selection operator (LASSO) algorithm. The correlation coefficients of all features extracted from preprocessed CE-T1WI images were compared using Fisher's Z transformation, with a correlation coefficient (r) of 0.8 or greater considered very strong. Only features with smaller mean correlation coefficients were retained. Subsequently, a t test was performed, with significance determined at p < 0.05. Subsequently, to further diminish the quantity of ultimate feature predictors, the least absolute shrinkage and selection operator (LASSO) binary logistic regression model was employed. The proportionate contribution of LASSO regression varied from 0 to 1, and the features with the least mean square error (MSE) estimated from the testing subset with 10-fold cross-validation were selected.
Our study incorporated 20 classification algorithms, encompassing logistic regression, support vector machine with linear kernels (SVMLINEAR), support vector machine with Gaussian kernels (SVMRBF), support vector machine with polynomial kernels (SVMPOLY), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), ensemble learning with LDA (ELDA), ensemble learning with QDA, naïve Bayes (BAYES), probabilistic neural network (PNN), artificial neural network (ANN), decision tree (DT), random forest (RF), grading boost with random undersampling boosting (GBRUSB), grading boost with AdaBoostM1 (GBABM1), grading boost with gentle boosting (GBGB), grading boost with logit boosting (GBLB), grading boost with robust boosting (GBRB), grading boost with LPBoosting (GBLPB), and grading boost with total boosting (GBTB). Patients were categorized as the training cohort (< 2018/7/1) and the prospective testing cohort (≥ 2018/7/1). The area under the curve (AUC) of each model was computed to evaluate their diagnostic performance.
Assessment of Predictive Models
The performance of these predictive models was verified in the validation group by receiver operating characteristic (ROC) curves. For training classic ML models, the cross-validation strategy will validate each training sample once and calculate a validation score for that sample. Thus, we calculated a cross-validation AUC by training a classic ML model. After predicting the independent test cohort, we also calculated a test AUC and its 95% confidence interval for each trained model using the bootstrap strategy of sampling with replacement. In addition, an assessment was made regarding the clinical utility of the models through the application of decision curve analysis (DCA). The determination of standardized net benefit (sNB) was derived from the decision curve, with the sNB value ranging from 0 to 1. Subsequently, the risk threshold obtained from the decision curve was utilized to categorize patients into low-risk and high-risk groups, and measures of sensitivity and specificity were frequently computed and employed as indicators of usefulness.
Statistical Analysis
Statistical analysis was conducted with MATLAB 2021b Student Version. The independent-sample t test was applied to compare the mean value of age between the training cohort and the validation cohort. The chi-squared test or Fisher’s exact test was applied to assess the significance of the categorical variables between different groups. Two-tailed p values less than 0.05 were regarded as statistically significant. Logistic regression with the LASSO penalty and the SVM model was implemented using Statistics and Machine Learning Toolbox in MATLAB 2021b Student Version.