Related Experiment Video
Updated: Sep 27, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
PET|MR Open Source Radiomics Models for Treatment Outcome Prediction in Recurrent GBM
Montserrat Carles1, Alejandro Mora-Rubio1, Tobias Fechter2
1Biomedical Imaging Research Group (GIBI230), Health Research Institute La Fe, 46026 Valencia, Spain.
Background/Objectives:
Glioblastoma (GBM) is a hard-to-treat cancer with a 5-year survival rate of 5.7% without significant improvement in the past decades. Personalised strategies based on improved tumour characterisation and treatment-outcome prediction could support treatment adaptation and improve therapeutic efficacy. Magnetic Resonance (MR) is the standard imaging modality for GBM radiotherapy (RT) planning. Positron Emission Tomography (PET) with O-(2)-[^18F]fluoroethyl-L-tyrosine (FET) is recommended for distinguishing recurrence from pseudo-progression. However, evidence on PET|MR complementarity for predicting treatment outcome in recurrent GBM remains limited.
Methods:
We evaluated FET-PET|MR biomarkers to predict Time-To-Progression (TTP), Overall Survival (OS) and Acute Recurrence (AR) in a prospective cohort of 185 recurrent GBM patients from 15 institutions. T1-weighted contrast-enhanced, Fluid-Attenuation Inversion-Recovery, Apparent-Diffusion-Coefficient maps and FET-PET images were analysed. Gross- and Planning-Target Volumes (GTV/PTV) were manually delineated and the intersection of PET|MR-GTVs was defined as MR∩PET. Radiomics models were developed (5-fold cross-validation) and evaluated in a held-out test set. Additionally, nnUNet was used to predict MR∩PET from MR sequences.
Results:
MR∩PET volume yielded the highest number of statistically significant models, 15 versus 6 for PET|MR-GTVs and 0 for PET|MR-PTVs. Of these 21 models, 19 required the inclusion of PET imaging. The best-performing models discriminated between short and long OS, with p < 0.0001 in validation and test, and predicted AR, with ROC-AUC(validation) = 0.81 and AUC(test) = 0.63. By combining both models, poor responders (short OS and/or AR) were identified with Sensitivity = 82% and Specificity = 75%. MR-based prediction of MR∩PET showed a Dice Similarity Coefficient (test) = 0.87 ± 0.12.
Conclusions:
The PET|MR radiomic models derived from our multicentre prospective cohort support the identification of recurrent GBM patients with poor RT outcomes, potentially enabling future personalised strategies for treatment improvement.
