This retrospective study included 125 patients with newly diagnosed GBM undergoing open craniotomy at the First Affiliated Hospital of Zhengzhou University from January 2018 to January 2020 according to the following criteria. The inclusion criteria were as follows: 1) patients with GBM confirmed by pathological report; 2) patients with complete data of medical and imaging records before surgery; and 3) patients with standard treatment, i.e., maximal safe surgical resection followed by radiotherapy plus adjuvant chemotherapy with temozolomide, then maintenance with 6–12 months of temozolomide as single-agent therapy. The exclusion criteria were as follows: 1) patients with biopsy only; 2) patients without complete medical records; 3) patients with incomplete image data and image artifacts; and 4) patients with radiotherapy or chemotherapy alone after surgery. The medical ethics committee of the First Affiliated Hospital of Zhengzhou University approved this retrospective study.

There were 4 clinical features and 2 conventional imaging features collected for each patient, including age at diagnosis, gender, preoperative Karnofsky performance status (pKPS), preoperative epilepsy status (pEPI), located lobe (frontal, temporal, parietal, occipital, insular, corpus callosum), and hemisphere (left, right, bilateral). Regular follow-up was applied to every patient until death or June 2021 through the clinic or phone, once every month for the first 6 months after surgery, and every 3–6 months thereafter. Each patient was separated into short or long survival group according to the OS of 12 months. Then, the patients were divided randomly into training cohort (n = 87, 70%) used for model construction and validation cohort (n = 38, 30%) used for model evaluation.

Four imaging sequences were selected from the head MRI undergone before surgery of each patient, i.e., T1-weighted contrast-enhanced imaging (T1C), T1-weighted imaging (T1), T2-weighted imaging (T2), and T2-weighted fluid-attenuated inversion recovery imaging (T2F). The imaging was performed on five models of MRI scanners from two manufacturers, i.e., Verio, Prisma, TrioTim, and Skyra of Siemens and Discovery MR750 of GE Medical Systems. The T1 sequence was acquired with the following range of parameters: repetition time (TR)/echo time (TE), 163–1,750.03 ms/2.46–25.176 ms; slice thickness, 5 mm; spacing between slices, 6.5–6.75 mm. The T1C sequence was acquired with the following range of parameters: TR/TE, 21–3,900 ms/2.32–92 ms; slice thickness, 0.9–5 mm; spacing between slices, 6.50–6.75 mm. The T2 sequence was acquired with the following range of parameters: TR/TE, 3,800–5,673.8213 ms/92–117 ms; slice thickness, 5 mm; spacing between slices, 6.5–6.75 mm. The T2F sequence was acquired with the following range of parameters: TR/TE, 5,000–8,400 ms/81–157.732 ms; slice thickness, 5 mm; spacing between slices, 6.5–6.75 mm.

Image preprocessing was performed by 3D Slicer software (v4.11.0). First, skull-stripping was executed by the Swiss Skull Stripper module, and T1, T2, and T2F sequence images were registered to T1C sequence images. Next, N4 bias field correction was applied to correct the intensity unevenness of each sequence. Ultimately, image normalization (normalizeScale = 100) and image resampling (ResamplePixelSpacing = [3, 3, 3]) were performed in Python environment by the PyRadiomics package.

The region of interest (ROI) was delineated on T1C using 3D Slicer software separately by two neurosurgeons, who were blinded to the clinical data and had over 5 years of clinical experience.

The radiomics feature extraction was executed in the Python environment with the PyRadiomics package, which was an open-source python package for the extraction of Radiomics features from medical imaging. The Supplementary Material Data S1 supplied the detail parameter settings of feature extraction. There were seven image types (Original, Wavelet, Laplacian of Gaussian (LoG sigma [3.0, 5.0]), Square, SquareRoot, Logarithm, Exponential) and six feature classes (shape, first-order statistics, gray-level co-occurrence matrix (glcm), gray-level run length matrix (glrlm), gray-level size zone matrix (glszm), gray-level dependence matrix (gldm)) adopted for each sequence.

Three feature selection steps were performed to avoid overfitting before establishing the model. First, the stable features, which were defined by intraclass correlation coefficients (ICCs) >0.8, were selected between the two groups of ROIs drawn by the two neurosurgeons (14). Next, the t-test was applied to each stable feature between short and long survival cases. Then, the least absolute shrinkage and selection operator (LASSO) regression algorithm was applied to analyze these features whose p-values were less than 0.05 in the t-test. The optimal λ value in LASSO was automatically selected by 10-fold cross-validation with a maximum area under the curve (AUC) criterion, where the final value of λ yielded the maximum AUC.

To understand the ability of these OS status-related radiomics features to discriminate the OS state, three supervised machine learning algorithms were applied in the training cohort, including random forest (RF) algorithm, support vector machine (SVM) algorithm, and logistic regression (LR) algorithm, and tested in the validation cohort. The performance of training and validation cohorts was evaluated by AUC. Then, these features with non-zero coefficients were used to construct the Radscore for each patient. Finally, the radiomics nomogram based on Radscore from the training cohort was established by logistic regression and assessed by the validation cohort. For the two cohorts, the discriminative ability of the radiomics nomogram was quantitatively measured using AUC, and the calibration curves were plotted based on observed probabilities and the nomogram-estimated probabilities (15). To evaluate the clinical utility of the radiomics nomogram, the decision curve analysis (DCA) was executed by calculating the net benefits at different threshold probabilities in the combined training and validation cohorts (16). The flowchart of this research is shown in Figure 1.

The continuous variables were analyzed by Student’s t-test or Mann–Whitney U-test, and chi-square test, Yates’ correction, or Fisher’s exact probabilities were performed in the categorical characteristics. Generally, a two-sided p-value < 0.05 was considered statistically significant. All data analyses were performed in Python (v3.7.6) and R software (v4.0.5).

A total of 320 newly diagnosed GBM patients were collected. According to the inclusion and exclusion criteria, we included 177 patients whose image quality met the criteria and 52 patients without standard treatment were excluded. 125 patients were finally included in our study (87 cases in the training cohort and 38 cases in the validation cohort). The clinical features, conventional imaging features, and Radscores of patients in the training cohort, validation cohort and total cohort, are summarized in Table 1. There were no significant differences between the short OS and long OS groups in age, gender, pEPI, pKPS, located lobe (frontal, temporal, parietal, occipital, insular, corpus callosum), and hemisphere (left, right, bilateral). However, it was worth noting that Radscore had significant differences in the short OS and long OS groups (p < 0.001).

In this study, a total of 1,304 radiomics features for each sequence were extracted, including 14 shape features, 18 first-order statistics features, 68 texture features, 172 LoG features, 688 wavelet features, 86 square features, 86 square-root features, 86 logarithm features, and 86 exponential features. A total of 5,216 radiomics features were calculated from four imaging sequences for each patient. Among them, 5,080 radiomics features were stable after being screened by ICCs. After that, 777 radiomics features were selected by t-test. Ultimately, the optimal regulation weight λ (λ = 0.029470517025518096) was determined for the LASSO algorithm, and 21 features with non-zero coefficients were selected for OS stratification of GBM patients. The detailed names and weights of the 25 radiomics features are shown in Table 2 and Figure 2. It could be seen that the T1C sequence had a greater impact on OS stratification.

There were three supervised machine learning algorithms models constructed to determine the ability of these OS status-related radiomics features to discriminate the OS stratification. The detail performances of the three models are shown in Table 3. The SVM model performed best among the three models. The AUC, sensitivity, accuracy, and F1 score were 0.75, 0.93, 0.71, and 0.72 in the validation cohort, respectively.

Then, the Radscore for each patient in training and validation cohorts was constructed for further analysis, which was calculated by multiplying each feature coefficient by the corresponding feature value and summing. The corresponding fitting formula is listed in Supplementary Material Data S2. Patients with long OS showed higher Radscores than patients with short OS in both the training and validation cohorts ( Figure 3). In the training cohort, the average values of Radscore were significant differences in the short OS and long OS groups (-0.167 vs. 0.128, p < 0.001). Similarly, the mean Radscore of long OS was 0.165, which was significantly higher than that of short OS (-0.207, p < 0.001) in the validation cohort.

To establish the radiomics nomogram, the logistic regression based on Radscore, clinical features, and conventional imaging features was applied to find independent predictors of OS stratification by univariate and multivariable logistic regression. The results of logistic regression are presented in Table 4, which demonstrated that only Radscore was the significant independent predictor for OS stratification.

Then, the radiomics nomogram was constructed according to the multivariable logistic regression ( Figure 4). The ROC curve, which was based on the probability of long OS according to the Radscore, was used to evaluate the sensitivity and specificity of the nomogram. The AUCs of the nomogram were 0.877 and 0.919 in the training and validation cohorts, respectively, which indicated favorable sensitivity and specificity ( Figure 5). The calibration curve of the proposed nomogram based on the training cohort was constructed, and a favorable calibration was confirmed in the validation cohort ( Figure 6). Moreover, the result of DCA showed that the nomogram to stratify the OS of GBM patients could yield clinical net benefits ( Figure 7).