top of page
  • Foto do escritorSNOLA

Glioblastoma and radiotherapy: A multicenter AI study for Survival Predictions from MRI (GRASP study)

Glioblastoma is the most aggressive primary brain cancer in adults. MRI plays a crucial role in diagnosis, treatment planning, and response assessment. However, by the time radiotherapy finishes, various interventions might confound survival predictions. The study aimed to use deep learning models on MRI images taken immediately after radiotherapy to predict 8-month survival in glioblastoma IDH-wildtype patients.


The data included 206 glioblastoma patients from 11 UK centers, diagnosed between March 2014 and February 2022. The models were trained on data from 158 retrospective patients and tested on 19 retrospective and 29 prospective patients. The deep learning model used T2-weighted and contrast-enhanced T1-weighted MRI images, combined with demographic and treatment data. Separate models were trained for imaging and non-imaging data, and a combined model was also evaluated. The imaging model outperformed the non-imaging model, with an area under the receiver-operating characteristic curve (AUC) of 0.93 compared to 0.79.


The patient cohort included 31.1% short-term survivors (surviving less than 8 months post-radiotherapy). Longer survival was associated with higher MGMT methylation percentage, non-deep-seated tumors, resection, and standard radiotherapy and temozolomide doses, as per Stupp protocol. Logistic regression with reduced features was the best -performing nonimaging model, using features like sex, MGMT status, surgery type, and treatment doses.


The study presents the first known model using MRI to distinguish between short-term and long-term survivors within 8 months post-radiotherapy. The model's accuracy was robust across different centers and test sets, demonstrating its generalizability. The study concludes that the deep learning model can serve as a reliable prognostic biomarker, identifying patients who may require early second-line treatments or clinical trial enrollment. The model’s integration into clinical practice could improve management decisions by providing accurate survival predictions from the first post-radiotherapy MRI.

Journal: Neuro-Oncology

26(6), 1138–1151, 2024 | | Advance Access date 29 January 2024

Author: Maria Thereza Mansur Starling

Mini-CV: Radiation Oncologist; MBA in Health Management from FGV-SP; Fellowship at London Regional Cancer Centre, Canada.


7 visualizações0 comentário


bottom of page