A Transformer-Based Deep Learning Model for Prediction of Temozolomide Resistance in Glioblastoma Using Pretreatment MRI Images.
Authors
Affiliations (10)
Affiliations (10)
- Department of Radiology, The Affiliated Hospital of Qingdao University, 16 Jiangsu Road, Qingdao 266003, Shandong, China (C.X., J.Z., M.L., X.L.).
- Department of Radiology, Sichuan Provincial People's Hospital, University of Electronic Science and Technology of China, Chengdu, China (S.L.).
- Department of Radiology, Lanzhou University Second Hospital, Cuiyingmen No. 82, Chengguan District, Lanzhou, China (Y.L., T.H., Q.Z., J.Z.).
- Department of Radiology, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China (M.J.).
- College of Electronic Information Engineering, Shandong University of Science and Technology, Qingdao, China (Q.L.).
- Department of Pathology, The Affiliated Hospital of Qingdao University, 16 Jiangsu Road, Qingdao, Shandong, China (W.F.).
- Center for Interventional Medicine, The Affiliated Hospital of Qingdao University, 16 Jiangsu Road, Qingdao, Shandong, China (P.H.).
- Department of Neurosurgery, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, China (C.L.).
- Department of Radiology, Second Affiliated Hospital of Navy Medical University, Shanghai, China (F.L.).
- Department of Radiology, The Affiliated Hospital of Qingdao University, 16 Jiangsu Road, Qingdao 266003, Shandong, China (C.X., J.Z., M.L., X.L.). Electronic address: [email protected].
Abstract
Glioblastoma (GBM) is a lethal tumor where temozolomide (TMZ) chemotherapy often faces resistance. We develop a deep learning model based on a Vision Transformer (ViT) architecture for pretreatment prediction of TMZ resistance. We retrospectively enrolled 314 GBM patients across four centers between January 2021 and December 2024. The cohort was divided into a training cohort, an internal validation cohort, and two external validation cohorts. Patients were categorized into resistant and sensitive groups based on the best radiological response after Stupp chemoradiotherapy. A ViT-based deep learning network (DLN) was developed by selecting the optimal peritumoral expansion margin. A fusion model was subsequently constructed by concatenating the DLN with MGMT methylation status. Model performance was evaluated using calibration and decision curve analysis (DCA). An expansion distance of 10 mm was selected as optimal for constructing the DLN model. The fusion model outperformed both the DLN and the MGMT in the training, internal validation, and external validation cohorts, demonstrating robust performance in predicting TMZ resistance with Area Under the Curve (AUC) values of 0.855 (95% CI, 0.802-0.905), 0.917 (95% CI, 0.737-1.000), 0.806 (95% CI, 0.632-0.937), and 0.848 (95% CI, 0.673-0.964), respectively. Decision curve analysis indicated that the fusion model provided a high net clinical benefit. The fusion model provides a precise and personalized risk assessment for TMZ resistance in GBM. This approach has the potential to spare resistant patients from unnecessary chemotherapy while identifying high-risk individuals who may benefit from more aggressive treatment strategies.