2696 - Predicting the Region of Tumor Microinfiltration within the Peritumoral Edema of Glioblastoma Using a Multiparametric Magnetic Resonance Model
Presenter(s)
J. Zhang1,2, H. Liu3, Y. Wu3, M. Du4, Y. Gao4, and B. Li1,5; 1Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Key Laboratory of Cancer Prevention and Therapy, Tianjing, China, 2Cancer Center, Linyi People`s Hospital, Shandong Second Medical University, Linyi, ??, China, 3School of Clinical Medicine, Shandong Second Medical University, WeiFang, China, 4Cancer Center, Linyi People`s Hospital, Shandong Second Medical University, Linyi, China, 5Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China
Purpose/Objective(s): Current research confirms that tumor microinfiltration (TMI) exists within the peritumoral edema (PTE) of glioblastoma (GBM). Predicting the extent of TMI within the PTE of GBM using Magnetic Resonance Imaging(MRI)examinations is clinically crucial for optimizing the extent of surgical resection and planning the target volume for postoperative radiotherapy.
Materials/Methods: This study cohort was divided into a training group (TG), a group validation (VG), and an image-pathology point-to-point VG (IMVG). The TG included 50 cases of GBM and 50 cases of brain metastasis (BM). The VG included 25 cases of GBM and 19 cases of meningioma (MG). The IMVG consisted of 10 additional GBM cases. Gray-level histogram parameters (GLHP) of peritumoral edema (PTE) were extracted from contrast-enhanced T1-weighted imaging (T1WI), T2 fluid-attenuated inversion recovery (T2FLAIR), and apparent diffusion coefficient (ADC) maps. Univariate analysis was performed on PTE parameters around different tumor lesions. LASSO regression was applied to remove collinear parameters. Subsequently, a forward stepwise logistic regression model (FSLRM), support vector machine (SVM), and random forest (RF) analysis were used to construct a prediction model for tumor microenvironment (TMI) within PTE. Model effectiveness was evaluated using receiver operating characteristic (ROC) curves, and the optimal threshold was explored. In IM VG, a TMI prediction map was generated using Python, based on the optimal threshold, and the coincidence rate between predicted TMI and biopsy pathology results was calculated.
Results: Univariate analysis in the TG showed significant differences in GLHP between PTE bands of GBM and BM lesions within 2 cm of tumor margins. After removing collinear parameters using LASSO regression, logistic regression models from ADC maps or contrast-enhanced T1WI sequences outperformed those from T2FLAIR sequences. The model pairing the 1-cm PTE band of GBM lesions with the 2-cm or 3-cm PTE band of BM lesions (PTE_GMB1cm_versus_BM2cm or PTE_GMB1cm_versus_BM3cm) showed better discrimination, with AUC values of 0.717 and 0.816 (ADC) and 0.785 and 0.804 (contrast-enhanced T1WI), respectively. Combining ADC and contrast-enhanced T1WI parameters into SVM and RF models revealed superior performance of the RF model. The RF model for PTE_GMB1cm_versus_BM2cm included six parameters: Ratio-maxi-enhancedT1/mean-enhancedT1, Ratio-maxiADC/meanADC, Ratio-miniADC/meanADC, skewness-enhancedT1, miniADC, and mini-enhancedT1, with an AUC of 0.836 and an optimal threshold of 0.71. In the VG, the AUC was 0.844. In the IMVG, the overall effective rate of predicting TMI within PTE using the RF model was 81.25%.
Conclusion: The multiparametric RF model based on the GLHP of the preoperative ADC map and contrast-enhanced T1WI sequence images can effectively predict the extent of TMI within PTE of GBM.