What question did this study set out to answer?

To develop a model that accurately predicts tumor micro-infiltration within peritumoral edema in glioblastoma using multiparametric MRI.

June 17, 2026Open Access

Predicting the region of tumor micro-infiltration in glioblastoma peritumoral edema using a multiparametric mri radiomics model

Key Points

To develop a model that accurately predicts tumor micro-infiltration within peritumoral edema in glioblastoma using multiparametric MRI.
Training group consisted of 50 glioblastoma and 50 brain metastasis tumor cases, with a validation group of 25 glioblastoma and 19 meningioma cases.
Utilized LASSO regression to remove collinear parameters and developed predictions using logistic regression, SVM, and random forest.
Efficiency of predictions evaluated through ROC curves, achieving an AUC of 0.836 in the training group and 0.844 in the validation group.
The random forest model with 6 parameters showed superior discrimination, with a TMI prediction model achieving overall accuracy of 81.25% in image-pathology point-to-point validation.
Univariate analysis revealed significant GLHP differences between glioblastoma and brain metastasis peritumoral edema within 3 cm.
The model incorporating ADC and CE-T1WI parameters outperformed others in predicting TMI.

Abstract

To predict tumor micro-infiltration (TMI) within glioblastoma (GBM) peritumoral edema (PTE) using MRI for postoperative radiation planning. Study cohort consisted of a training group (TG) with 50 GBM and 50 brain metastasis tumor cases, a validation group (VG) with 25 GBM and 19 meningioma cases, and an image-pathology point-to-point VG (IPVG) with 10 additional GBM cases. Besides univariate analysis of gray-level histogram parameters (GLHP) of peritumoral edema (PTE) from contrast-enhanced T1-weighted imaging (CE-T1WI), T2 fluid-attenuated inversion recovery (T2FLAIR), and apparent diffusion coefficient (ADC) maps, LASSO regression removed collinear parameters. Then, forward stepwise logistic regression, SVM, and random forest were used to build TMI prediction models in PTE, with efficiency evaluated via ROC curves. In the IPVG, the coincidence rate between Python-based TMI predictions and biopsy pathology results was calculated. Univariate analysis in the TG revealed that GLHP differences between GBM and BM PTE belts were prominent within 3 cm of PTE. After removing collinear parameters, the ADC map or CE-T1WI-based prediction model outperformed the T2FLAIR-based one. The model with the 1-cm GBM PTE belt and 2-cm BM PTE belt pairing showed superior discrimination. Incorporating ADC and CE-T1WI parameters, the RF model with 6 parameters (Ratio-maxi-enhancedT1/mean-enhancedT1, Ratio-miniADC/meanADC, etc.) demonstrated the best discrimination than other models, with an ROC curve (AUC) of 0.836 in the TG and 0.844 in the VG. In the IPVG, the TMI prediction model had an overall accuracy of 81.25%. The multiparametric RF model preoperative MRI can predict the TMI within PTE of GBM.

Ask AI

Mark Helpful

Bookmark

Relay

View Full Paper