Integrating Intratumoral and Peritumoral Features via a Cross-Attention Transformer on Super-resolution MRI for Glioblastoma Survival Risk Stratification: A Multicenter Study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Tianjin Fourth Central Hospital, The Affiliated Hospital of Tianjin Medical University, Tianjin 300140, China (E.Z., X.Z., T.F., Y.W., X.G., J.L.).
- Department of Radiology, Tianjin Union Medical Center, The First Affiliated Hospital of Nankai University, Tianjin 300121, China (X.L., J.L.).
- Tianjin Fourth Central Hospital, Tianjin University, Tianjin 300072, China (T.Z., J.L.).
- Department of Neurology, Tianjin Medical University General Hospital, Tianjin, China (Y.S.).
- Department of Radiology, The Second Affiliated Hospital of Dalian Medical University, Dalian, China (C.Y.).
- Department of Radiology, Tianjin Fourth Central Hospital, The Affiliated Hospital of Tianjin Medical University, Tianjin 300140, China (E.Z., X.Z., T.F., Y.W., X.G., J.L.); Department of Radiology, Tianjin Union Medical Center, The First Affiliated Hospital of Nankai University, Tianjin 300121, China (X.L., J.L.); Tianjin Fourth Central Hospital, Tianjin University, Tianjin 300072, China (T.Z., J.L.). Electronic address: [email protected].
Abstract
To develop and validate a Cross-Region Transformer using super-resolution (SR) magnetic resonance imaging (MRI) to integrate intratumoral and peritumoral features for overall survival (OS) prediction in glioblastoma. This multicenter study included 362 patients (training, n = 228; internal test, n = 57; external validation, n = 77). Preoperative contrast-enhanced T1-weighted and T2-weighted MRI underwent generative adversarial network-based SR reconstruction. Intratumoral and 10-mm peritumoral features were extracted using 2.5D convolutional neural networks and fused by asymmetric cross-attention to generate a deep learning-based risk score (DLRS), which was combined with independent clinical factors. Performance was evaluated using C-index, time-dependent receiver operating characteristic analysis, calibration, decision curve analysis, Gradient-weighted Class Activation Mapping, and SHapley Additive exPlanations. The Cross-Region Transformer achieved C-indices of 0.787 (95% confidence interval [CI], 0.734-0.840), 0.711 (95% CI, 0.593-0.829), and 0.726 (95% CI, 0.627-0.826) in the training, internal test, and external validation cohorts, respectively, with numerically higher discrimination than single-region and feature-concatenation models. The combined clinical-DLRS model, incorporating the Stupp regimen, number of lesions, Ki-67 index, and DLRS, achieved corresponding C-indices of 0.836 (95% CI, 0.788-0.884), 0.742 (95% CI, 0.628-0.855), and 0.754 (95% CI, 0.658-0.850). The DLRS significantly stratified OS in external validation (median, 7.5 vs 23.5 months; P < 0.001). Interpretability analyses highlighted enhancing margins within the necrotic core and the intratumoral-peritumoral interface. The Cross-Region Transformer integrates complementary intratumoral and peritumoral representations from SR MRI. The resulting interpretable DLRS supports noninvasive survival prediction and risk stratification in glioblastoma.