.png){kind=icon}
.png){kind=icon}
Return
Multimodal PET–MR segmentation for glioblastoma: complementarity for treatment planning and recurrence definition
.png){kind=icon}
OA
AI
A
D
C
T
M
M
I
R
J
P
M
D
L
A
I
M
DOI：10.1186/s40644-026-01086-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In patients with Glioblastoma (GBM), Magnetic Resonance (MR) is used for tumour diagnosis and treatment planning. Positron Emission Tomography (PET) with O-(2)-18 F-Fluoroethyl-L-Tyrosine (FET) has been recommended to distinguish local recurrence from radiogenic alterations. However, clinical practice remains hindered by the time and expertise required for tumour and organs-at-risk (OARs) segmentation and the limited evidence of the added value of PET and its restricted availability across clinical centres. This study presents automatic segmentation models and a comprehensive evaluation of PET/MR complementary biological information for recurrent disease definition. The nnU-Net was employed for segmentation using manually defined contours on 1,610 patients from 33 institutions. Model performance was evaluated by Dice-Sørensen-Coefficient (DSC). PET/MR recurrence complementarity was evaluated in 185 patients by Wilcoxon-Signed-Rank test (WSRT), DSC and radiomic features (RF). In RF analysis, MR-Enhancing subregions were classified as MR∩PET or MR-Only, based on overlap with PET. For MR-RF showing significant MR∩PET/MR-Only differences (WSRT), discrimination was further assessed by classifying RF for 3 × 3 × 3-voxel subregions within MR-Enhancing in two volumes (greater/less than the RF cohort median) and evaluating Positive-Predictive-Value and Sensitivity with MR∩PET/MR-Only. Models’ performance in the test set resulted in DSC(MR-Enhancing) = 0.76 ± 0.24, DSC(MR-Edema) = 0.69 ± 0.23, DSC(PET-Uptake) = 0.71 ± 0.20, DSC(Planning-Target-Volume) = 0.93 ± 0.05, DSC(OARs) = 0.70 ± 0.13. Manual MR and FET-PET based GBM recurrence delineations differed significantly in size (p = 0.0497) and location (DSC = 0.45 ± 0.20). From the 37 MR-RF showing significant differences between MR∩PET and MR-Only (p < 0.05), none of the MR-Enhancing based RF-maps allowed spatial identification of PET findings (positive-predictive-value and sensitivity < 0.6). The resulted segmentation models could facilitate PET/MR integration in GBM treatment. PET/MR comparison supports the complementarity of FET-PET.
Keywords:
Glioblastoma
Convolutional neural network
Segmentation
Radiomics
Radiation oncology
Journal
IF:
3.5
Papers:
1.3K
Citations:
3.5K
Organization
Cited Papers