Moderate evidence
More
Brain scan AI split glioblastoma patients into 7.5 vs 23.5 months median survival in a 362-patient study
Study title: Integrating Intratumoral and Peritumoral Features via a Cross-Attention Transformer on Super-resolution MRI for Glioblastoma Survival Risk Stratification: A Multicenter Study
- Evidence type
- Controlled or validated comparison
- Sample
- n = 362
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-3070cite or check it
Study title: Integrating Intratumoral and Peritumoral Features via a Cross-Attention Transformer on Super-resolution MRI for Glioblastoma Survival Risk Stratification: A Multicenter Study
A multicenter study of 362 glioblastoma patients built a deep learning model, the Cross-Region Transformer, that reads preoperative MRI of the tumor and the 10-mm tissue around it to predict overall survival. In the external validation group of 77 patients, its risk score split patients into groups with median survival of 7.5 versus 23.5 months.
Read the study abstract
Rationale and objectives 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. Materials and methods 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. Intrat…
The gain
In a multicenter study of 362 glioblastoma patients, an AI risk score from super-resolution MRI separated external-validation patients into groups with median survival of 7.5 versus 23.5 months.
In brief
The model reached C-indices of 0.787 in training, 0.711 in the internal test group and 0.726 in the external validation group. Adding clinical factors (the Stupp regimen, number of lesions and Ki-67 index) raised these to 0.836, 0.742 and 0.754.
The work points to a noninvasive way to estimate survival risk from routine scans, though discrimination dropped from training to the test groups and the confidence intervals in the smaller cohorts are wide.
Main points
- The study included 362 patients: 228 for training, 57 for internal testing and 77 for external validation.
- Preoperative contrast-enhanced T1-weighted and T2-weighted MRI was enhanced with generative adversarial network-based super-resolution reconstruction.
- Features from the tumor and a 10-mm surrounding zone were extracted with 2.5D convolutional neural networks and fused by asymmetric cross-attention.
- The imaging model scored C-indices of 0.787, 0.711 and 0.726 in training, internal test and external validation.
- The combined clinical and imaging model scored 0.836, 0.742 and 0.754 in the same three groups.
- The imaging model had numerically higher discrimination than single-region and feature-concatenation models.
The full story
Researchers trained the model on 228 patients and tested it on 57 internal and 77 external patients. Images were first sharpened by super-resolution reconstruction, then tumor and peritumoral features were pulled out by convolutional networks and merged using cross-attention to produce a deep learning risk score.
The imaging-only score had a C-index of 0.726 (95% CI 0.627-0.826) in external validation. The combined model, which added the Stupp regimen, number of lesions and Ki-67 index, reached 0.754 (95% CI 0.658-0.850).
Performance was checked with time-dependent ROC analysis, calibration and decision curve analysis, and interpretability was examined with Grad-CAM and SHAP. The authors conclude the risk score supports noninvasive survival prediction and risk stratification.
Sources
1 on file- Academic Radiologydoi:10.1016/j.acra.2026.10.014
Talk it over
In the external validation group of 77 patients, the risk score split people into median survival of 7.5 versus 23.5 months. How would you weigh that split against the drop in C-index from 0.787 in training to 0.726?
Brain scan AI split glioblastoma patients into 7.5 vs 23.5 months median survival in a 362-patient studyhttps://truvace.com/item/integrating-intratumoral-and-peritumoral-features-via-a-cross-attention-transfor
In the external validation group of 77 patients, the risk score split people into median survival of 7.5 versus 23.5 months. How would you weigh that split against the drop in C-index from 0.787 in training to 0.726?
Answer in the threadThe model was trained on 228 patients and tested on 57 internal and 77 external patients, with wide confidence intervals in the smaller groups. What can groups that size tell us about how it would work in new hospitals?
Answer in the threadIf you were a patient with glioblastoma offered a risk score from your preoperative MRI, what would you want to know about how it was built before deciding how much weight to give it?
Answer in the thread
If you have more time
Adding the Stupp regimen, number of lesions and Ki-67 index raised the external C-index from 0.726 to 0.754. What would you ask clinicians about whether that gain is large enough to change care?
Answer in the threadWhat result from a future study, such as one following new patients over time, would settle whether this risk score should be used with real patients?
Answer in the thread
ace
What readers say