This prospective observation study was approved by the institutional ethics committee, and written consent were signed by the patients or legal guardians.
Patient Registration
114 untreated BSCM patients who was admitted at the outpatient department between January 2009 to December 2013 were enrolled in the prospective study (Chinese Clinical Trial Registry, prospective clinical trial registration no. ChiCTR-POC-17011575, http://www.chictr.org.cn/). The inclusion criteria: 1) definite BSCM diagnosis with the multiparametric MRI (T2-weighted imaging [T2WI], T1- weighted imaging [T1WI], contrast-enhanced T1- weighted imaging [CE-T1WI], and the susceptibility weighted imaging [SWI]). Exclusion criteria: 1) pregnancy and breastfeeding, 2) Prior surgery or radiotherapy, 3) severe complications that may affect the safe of patients or the evaluation accuracy, 4) lost in follow-up, and 5) surgery or radiotherapy immediately after the enrollment.
Clinical information
The patients’ clinical information was stored in the electronic medical system.
Prior hemorrhage was defined by the definite medical history and the relevant CT/MRI evidence. As focal neurological deficits (FNDs), cranial nerve palsy, long-tract deficits (motor and sensory abnormalities), and/or extrapyramidal symptoms specifically related to the anatomical location of the lesion were identified. To assess neurological function in each patient, two clinicians blinded to imaging data and hemorrhagic information used the modified Rankin Scale (mRS). To check for any new bleeding or absorption, in patients with MRI scans obtained more than 2 weeks before the visit, a new scan was requested. According to the latest T1WI and T2WI at the time of diagnosis, hemorrhage was defined radiologically as an intralesional or extra-lesional hemorrhage. Two neuroradiologists independently reviewed the radiological data.
A number of characteristics of the lesion were examined, including the location [13], the side, the depth [14], perilesional edema, DVA, equivalent size, the crossing of axial midpoint [10], and Zabramski type [15]. Three grades of the crossing of axial midpoint: I) none; II) mild, as a result of mass effect, the axial midpoint was crossed, and the lesion was predominantly unilateral; and III) severe, the lesion crossed the axial midpoint and was located medially. Based on the lesion-equivalent diameter (ABC)1/3, the lesion size was calculated. Depending on the thickness of the parenchyma between the boundary of the nidus and the pial/ependymal membrane, the depth of the lesion was determined: I) superficial, nidus that were exophytic or abutting the pial/ependymal membrane with a thickness < 1 mm, II) deep, the thickness ≥ 2 mm, and III) moderate, the thickness ≥ 1 and < 2 mm[4].
Follow-Up
Inception of the follow-up was the first clinic visit at our facility. After that, clinical evaluations were conducted at 3, 6, and 12 months and then annually thereafter or whenever new symptoms or worsening symptoms appeared. And, the same schedule was followed for MRI scans. A record of subsequent treatments, treatment effects, and any new hemorrhages was kept. When the patient has the clinical symptoms related to the lesion including the FND (acute, subacute, new, and worsening) or the severe headache and the radiological change including the overt hemorrhage inside or outside the previous lesion or the enlarge signal density change, the patient will be confirmed that the prospective hemorrhage[16]. In addition, after the appearance of symptoms, in terms of the time course of the hemorrhage, there are three categories of hemorrhages on MRI: hyperacute (24 hours, iso- or hyperintense on T1WI and iso- or hyperintense on T2WI), acute (1–3 days, iso- or hypointense on T1WI and hypointense on T2WI), or subacute (3–14 days, hyperintense on T1WI and hypo- or hyperintense on T2WI) [16]. Observations were censored when the first prospective hemorrhage occurred. HFS was set as the interval between the initial hemorrhage and the second hemorrhage.
Brain MRI Sequence
Figure 1 showed the workflow of this study. The axial T2WI and axial CE-T1WI were used in the study. Before the examination, all patients signed the informed consent and removed all metal objects. And, the examination was done on the spine position with the 3.0-T scanner (Siemens Magnetom Skyra). For T2WI and T1WI Repetition time was 4900ms and 1770ms; Echo time was 117.12ms and 9.4ms; Acquisition matrix was 320 × 288 and 256× 198; Flip angle was 90° and 150°, respectively. In addition, Field of view, slice thickness, and spacing between slices were both 100 × 100, 5 mm, and 6 mm. Furthermore, The CE-T1WI was done after the patients received the injection of the contrast agent (gadolinium-DTPA, Magnevist, 0.1 mmol/kg) with the parameter of T1WI. These images were stored as the DICOM format on the picture archiving and communication system.
Regions of Interest Delineating
Three dimensional regions of interest (ROI) delineating was done by a neuroradiologist with 8 years of experience with the help of MRIcron (http://www.mricro.com; University of South Carolina, Columbia, SC, USA) according to the edge of the hemorrhage/hemosiderin, and the associated DVAs were incorporated into the segmentation. Then, another neuroradiologist with 10 years of experience confirmed the results. Any disagreement between the two neuroradiologists will be solved by the neuroradiologist with 25 years of experience.
Radiomic Feature Extraction
PyRadiomics (version 2.1.2) was used to extract radiomic properties from ROIs. The detailed algorithms and feature explanations could be found at https://github.com/Radiomics/pyradiomics [17]. Gray values were discretized using a fixed bin width. The features of T2WI and CE-T1WI were extracted, respectively. Total 1,304 radiomic features were extracted from each ROI with six built-in filters including wavelet, Laplacian of Gaussian (LoG), square, square root, logarithm, and exponential. Finally, the features could be assigned into four categories that 14 first order statistics, 126 shape features, 688 wavelet features, and 476 texture classes[18].
In addition, the shape features primarily describe the size and shape of the BSCM region in three dimensions; based on commonly used and basic metrics, first-order statistics describe how intensities are distributed within the image region defined by the mask; texture features describing patterns or the spatial distribution of voxel intensities, which has four types that were gray level dependence matrix (GLDM), gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and gray level size zone (GLSZM); wavelet transform effectively separates textural information from high- and low-frequency components of the original image, which is similar with the Fourier transform analysis.
Radiomic Features Selection and radiomics signature building
All feature values were normalized based on minimum-maximum normalization with the following formula, and the Xmin and Xmin was the respective minimum and maximum values of the feature Xn:
$$ \text{N}\text{o}\text{r}\text{m}\text{a}\text{l}\text{i}\text{z}\text{e}\text{d} \text{X}\text{n}=\frac{\text{X}\text{n}-\text{X}\text{m}\text{i}\text{n}}{\text{X}\text{m}\text{a}\text{x}-\text{X}\text{n}}$$
Based on these normalized features, features were selected and models were trained. Moreover, since the radiomic features are numerous and complex, we needed to perform a selection process to reduce overfitting. On the training set, the selection process was conducted as described previously [19]. Three methods were used to prioritize the features: (I) univariate analysis was performed for each feature. Features with P < 0.1 were considered to be associated with HFS potentially and were selected into the following process[20]; (II) the Pearson correlation coefficients (hereafter denoted r) between each pair of features were computed to analyze the linear correlation. The features were divided into different groups to ensure all pairs of features in a group had a |r| greater than 0.8. To remove the redundant features, only the most important prognostic feature of each group (i.e. the feature that yielded the lowest P value in univariate analysis) was remained; (III) the elastic net approach [21] was used to select the most insightful features. Elastic net was a selection operator combining Least absolute shrinkage and selection operator and ridge regression. With tuning parameter alpha (0–1, step 0.1), the E-net was trained, and the λ was confirmed with 10-fold cross-validation, which followed the minimum standard deviation criteria. A model with final alpha and λ values was used to select features with non-zero coefficients. The goal is to rank the features and recursively remove the ones that contribute least to classification. Weights are assigned to the features in order to train the estimator. The smallest weight features will be pruned from the current set of features, and the procedure will be repeated until the desired number of features is reached. The linear combination of final features and their corresponding coefficients was used to calculated the radiomics score of each patient.
Model construction and evaluation
According to the Rad-score threshold calculated by the X title software, patients were stratified into high- and low-risk groups. A comparison of survival curves between high-risk and low-risk groups was analyzed in a training cohort and validated in a validation cohort. To determine the potential risk factors, univariate Cox regression analyses were performed in the training cohort. Next, a radiomics nomogram was constructed to predict postoperative survival based on the radiomics signature and potential risk factors. Using the C-index, the nomogram was evaluated for discrimination ability. The ROC curve was used to measure the prediction value of the model. Calibrating performance was measured by the calibration curve, which described the agreement between predicted survival probability and observed survival probability. In the whole cohort, DCA [22] calculated the net benefits at different threshold probabilities to assess the clinical value of the nomogram.
Statistical analysis
All Analyses were performed on R (version 4.2.1, R Foundation for Statistical Computing, Vienna, Austria) and X-tile (version 3.6.1, Yale university, http://medicine.yale.edu/lab/rimm/research/software.aspx) [23]. Normally distributed variables are presented as mean + standard deviation, while non-normally distributed variables are presented as median (interquartile range). A t-test or Mann-Whitney U test was used to compare differences between the two groups. Categorical variables were also expressed in frequency (percentage) and assessed using Pearson's chi-square test or Fisher' s exact test. The Kaplan-Meier method was used to calculate survival curves, and log-rank tests were used to compare them. For univariate analyses, the Cox regression analysis was used. Elastic net regression was analyzed with the glmnet package. DCA was carried out using the dca. R package, while the nomogram and calibration curve were established with the rms package. A ROC analysis was performed using the time ROC R package. Based on the C-index, the nomograms were evaluated for their predictive performance. Statistical significance was determined by a P value of 0.05.