We recruited patients who experienced treatment failure following standard STUPP regimens (Concurrent oral administration of TMZ 75 mg/m2 and at least 6 cycles of adjuvant chemotherapy during radiotherapy) for rGM between March 2018 and February 2022 at the Radiotherapy Center of the First Affiliated Hospital of Nanjing Medical University. Diagnosis of glioma recurrence was based on the expertise of a neurologist, radiologist, and radiotherapist. The inclusion criteria for our study were: (1) Age of 18 years or older; (2) Karnofsky Performance Status (KPS) score of  ≥ 60; (3) Pathologically confirmed diagnosis of malignant gliomas according to WHO II, III, or IV grading; (4) Presence of recurrence or residual lesions confirmed by MRI following treatment with standard STUPP regimens; (5) followed by the Response Assessment in Neuro-Oncology(RANO) criteria, which require the presence of at least one measurable lesion for accurate identification of disease progression, pseudoprogression, or radiation necrosis[25]; (6)No use of other targeted drugs during treatment; (7) No functional impairment of important organs, no other contraindications to treatment. The exclusion criteria were as follows: (1) Patients with serious underlying systemic diseases or a life expectancy of less than 3 months; (2) Incomplete clinical data that prevents effective evaluation of clinical efficacy; (3) A history of previous mental illness; and (4) Presence of other malignant tumors. The flowchart is presented in Fig. 1.

The flowchart of this study

Patients in our study received oral administration of anlotinib once daily for 14 days every 21 days. For patients receiving intensified TMZ therapy, TMZ was administered at a dose of 100 mg/m2 once daily for 7 days every 14 days. Standard dosage regimens consisted of 150–200 mg/m2 orally for 5 days every 28 days until progression, intolerable toxicity, or death. The initial dosage of anlotinib was 12 mg. If grade 2 hemorrhage or other grade 3 or 4 adverse events occurred, the dose could be reduced to 10 or 8 mg.

Patients suspected of recurrence approximately 3 months after treatment underwent dynamic observation before undergoing MRI and functional MR imaging. The assessment of recurrence criteria was mainly carried out by MDT (consisting of experts from Radiotherapy Department, Neurosurgery Department, and Radiology Department), who followed the RANO standard and comprehensively analyzed the clinical manifestations, enhancement scope, and time of occurrence of each patient. All patients underwent MRI, and during the MDT process, we evaluated whether patients needed additional functional MRI (MRS and PWI) to assist with diagnosis. On conventional MRI, the signal pattern of recurrent glioma exhibited an apparent lack of uniformity, displaying a mixture of T1/T2 signal shadows along with the presence of distinct peripheral finger edema shadows. The enhanced scan further revealed prominent abnormal enhancement in both annular and nodular forms [25]. Most recurrent gliomas appear as high perfusion on Perfusion weighted imaging (PWI) [26, 27], and the sensitivity for diagnosing glioma recurrence is notably higher when the Cho/NAA ratio in Magnetic resonance spectrum (MRS) is greater than 1.8 [28]. The primary endpoints being the objective response rate (ORR), defined as the proportion of patients with complete or partial responses and stable disease for at least 4 weeks, and the disease control rate (DCR), defined as the proportion of patients with complete or partial responses and stable disease for at least 4 weeks. Secondary endpoints included PFS and OS.

In this study, image acquisition was performed using a 3.0 T magnetic resonance imaging scanner (Siemens MAGNETOM Vida 3.0 T MR) equipped with a cranial 8-channel orthogonal coil. All enrolled patients received routine T1-Weighted Imaging(T1WI) and T2-Fluid Attenuated Inversion Recovery(T2FLAIR) sequence scanning on transverse cephalic scan, followed by intravenous injection of contrast agent Gd⁃DTPA (0.1 mmol/kg body weight) for T1WI axial and sagittal thin layer 3D T1-weighted contrast-enhanced(T1C) scanning, coronal enhancement was reconstructed by sagittal thin layer enhancement post-processing. Conventional scan parameters were utilized with the following specifications: T1WI: TR 400 ms, TE 2.48 ms, layer thickness 5 mm, matrix 320 × 256, FOV 230 mm × 230 mm; Sagittal position thin layer 3D T1C: TR 1600 ms, TE 1.8 ms, layer thickness 1 mm, voxel size 0.7 × 0.7 × 1.0mm3, matrix 320 × 288, FOV 220 mm × 220 mm; T2FLAIR: TR 8000 ms, TE 97 ms, layer thickness 5 mm, layer spacing 1 mm, TI 2300 ms, Matrix 256 × 256, FOV 230 mm × 230 mm.

Two radiologists independently performed MRI feature analysis. The readers, who were blinded to the clinical-pathological data, segmented the volume of interest (VOI) using ITK-SNAP software (version 4.0; http://www.itksnap.org/pmwiki/pmwiki.php).The outermost boundaries of the tumor and edema were delineated artificially on the T1C and T2FLAIR images, respectively. If there are differences in opinions between the two radiologists, a third expert specialized in the field of gliomas was consulted. To account for variations in the MRI scanner protocols, several pre-processing steps were applied. The first step was to reorient the images and labels such that they had a consistent anatomical orientation with RAS axis codes. The next step involved resampling the images and labels using bilinear interpolation for the images and nearest neighbor interpolation for the labels to ensure a voxel spacing of 1 × 1x1 mm3. Then, the intensity values of the images were rescaled to have a range of [0,1] using a linear transformation, with optional clipping. Finally, removal of the background region of the images and labels was performed based on the foreground mask computed from the source image.

All radiomics signatures were extracted using Pyradiomic's (version 3.0.1; http://pyradiomics.readthedocs.io) in-house feature analysis procedure (Supplementary document). For each MRI sequence, we extracted 7 feature groups, shape (28 features), first-order statistical features (720 features), NGTDM (200 features), GLCM (880 features), GLDM (560 features), GLRLM (640 features) and GLSZM (640 features).

Firstly, characteristics with stability and repeatability were selected using Spearman's rank correlation coefficient test. Next, all features were standardized using the Z-score method, and hypothesis testing was employed to screen out features with significant differences between the two groups to ensure the validity of all features. Feature dimensionality reduction was achieved using Least Absolute Shrinkage And Selection Operator (LASSO) Regression, which can effectively reduce overfitting and improve prediction accuracy. The regularization parameter (λ) in feature selection was adjusted by tenfold cross-validation of the minimum criteria. Radiomics signatures were developed using selected features with nonzero coefficients. To build a predictive model of radiomics in a supervised learning manner, Support Vector Logistic Regression (LR), Machine (SVM), K-Nearest Neighbor (KNN), Extremely Randomized Trees (Extra Trees), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Multilayer Perceptron (MLP) classifiers were utilized in the training cohort. fivefold cross verification was used to select the best performance model. Firstly, multiple model choices are verified on the training set and testing set, and the model with minimum average error is selected. After selecting the appropriate model, the training set can be combined with the testing set, and the model can be trained again on it to get the final model, and then the test set can be used to test its generalization ability. The radiomic score (Rad-score) was calculated by a linear combination of selected features weighted by their respective coefficients. Model evaluation was performed using Area Under the Curve (AUC) values, accuracy, specificity, and sensitivity. Finally, the application value of the final model was analyzed through Decision Curve Analysis (DCA).Fig. 2 provides an illustration of the image analysis workflow in our study.

Workflow of radiomic analysis of this study

Statistical analyses in this study were performed using IBM SPSS Statistics software(version 25.0, https://www.ibm.com) and R statistical software(version 3.3.3, https://www.r-project.org). Survival analyses were conducted using the Kaplan–Meier method with a 95% confidence interval and compared using the Breslow test. The features were analyzed using Mann–Whitney U-test, t-test, or χ2 test, with statistical significance set at p < 0.05.For features with high repeatability, Spearman's rank correlation coefficient was also applied to calculate the correlation between features, and features with a correlation coefficient greater than 0.9 between any two features were retained. To maximize the ability to describe features, a greedy recursive deletion strategy was utilized for filtering features, which involved deleting the most redundant features in the current set each time. LASSO regression modeling was performed using the Python scikit-learn package.