Related Experiment Video
Updated: May 10, 2026
Glioblastoma Relapse Post-Resection Model for Therapeutic Hydrogel Investigations
Published on: February 24, 2023
Transformer-Based Deep Learning Model for Predicting Recurrence in High-Grade Glioma
Xin Wang1, Mingjun Ding1, Dan Zong1
1Department of Radiotherapy, The Affiliated Cancer Hospital of Nanjing Medical University, Jiangsu Cancer Hospital, Jiangsu Institute of Cancer Research, Nanjing, China.
Introduction:
The first year after treatment for high-grade glioma (HGG) is recognized as the peak interval for recurrence. Accurate prediction of recurrence during this period is critical for timely management and early intervention. This study aimed to develop a fusion model that integrates MRI-derived features with clinical variables.
Methods:
A retrospective analysis was conducted on 309 postoperative patients with HGG who received intensity-modulated radiation therapy (IMRT) at Jiangsu Cancer Hospital from 2016 to 2023. Patients were randomly assigned (7:3) to training (n = 216) and test (n = 93) cohorts. Clinical variables were screened using univariate analyses. Regions of interest (ROIs)-including the postoperative cavity and peritumoral edema-were manually segmented on T2-weighted images by two senior physicians using ITK-SNAP. Pretrained DenseNet-121, DenseNet-201, and DenseNet-169 architectures were used to extract deep-learning features. These features were used to construct deep-learning models; all models were trained using fivefold cross-validation on the training cohort. Gradient-weighted class activation maps (Grad-CAM) were generated to visualize model attention. Model performance was evaluated using receiver operating characteristic (ROC) analysis, calibration curves, and decision curve analysis (DCA).
Results:
Of the 309 patients, 134 experienced recurrence within one year. Integrating clinical variables with transformer-based models yielded substantial performance gains. Among the integrated models, Combined-Transformer-DenseNet121 achieved the best performance, with an AUC of 0.903 (95% CI, 0.862-0.943) in the training cohort and 0.747 (95% CI, 0.642-0.852) in the test cohort. Calibration fidelity was improved; Hosmer-Lemeshow tests indicated excellent agreement between predicted and observed outcomes (HL statistic, 0.166 [training] vs. 0.158 [test]). DCA demonstrated superior clinical utility in both training and test cohorts, consistently yielding the highest net benefit across threshold probabilities when model-predicted probabilities were applied.
Conclusions:
The proposed model demonstrated superior performance for predicting one-year recurrence in high-grade glioma compared with traditional approaches, offering high accuracy and facilitating early identification of high-risk patients.