RPS 2111 - Gliomas: advanced techniques in neuro-oncology imaging

Author Block: A. h. Shah, S. Sen, A. Chandra, A. Gehani; Kolkata/IN
Purpose: RANO 2.0 refines trial-grade response criteria but remains complex and inconsistently applied in routine practice. The Brain Tumor Reporting and Data System (BT-RADS) offers a structured lexicon for follow-up, yet head-to-head outcome-based validation against RANO 2.0 is lacking. We compared diagnostic accuracy, reproducibility, survival stratification, and downstream management impact of BT-RADS versus RANO 2.0 in treated gliomas.
Methods or Background: We retrospectively reviewed 536 patients with treated gliomas (2012–2024; 1,082 follow-up MRIs). Two neuroradiologists independently applied both BT-RADS and RANO 2.0. Truth standards were histopathology, MDT consensus, or ≥12-month follow-up. Endpoints included diagnostic performance, interobserver agreement (κ), Kaplan-Meier progression-free survival (PFS) and overall survival (OS), and concordance with MDT management decisions. Published BT-RADS work confirms feasibility, but survival-based, RANO-anchored validation has not been reported.
Results or Findings: BT-RADS achieved higher diagnostic accuracy (88%) than RANO 2.0 (77%), with better reproducibility (κ = 0.80 vs 0.61). Survival separated cleanly by BT-RADS: median PFS 12 months / OS 22 months for BT-RADS 3, versus PFS 6 months / OS 11 months for BT-RADS 4 (log-rank p<0.001). RANO 2.0 categories overlapped substantially, limiting prognostic clarity. BT-RADS also reduced false positives from pseudoprogression, operationalizing RANO’s “confirm on repeat scan” rule. Clinical impact was evident: MDT escalation occurred in 78% of BT-RADS 4 and observation in 85% of BT-RADS 2, higher alignment than with RANO 2.0.
Conclusion: In real-world follow-up, BT-RADS is not only more reproducible but also more clinically meaningful than RANO 2.0-cleanly stratifying survival and aligning with actual MDT decisions. This is the first survival-stratified, decision-linked validation of BT-RADS versus RANO, supporting structured lexicons in daily glioma care.
Limitations: Single-center retrospective design; heterogeneous treatment regimen.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: A. M. Viveros¹, C. Majós¹, P. Naval-Baudin¹, J. I. García García¹, M. Cos Domingo¹, L. Oleaga¹, M. Smits², A. Pons Escoda¹; ¹Barcelona/ES, ²Rotterdam/NL
Purpose: This study aims to assess the diagnostic performance of VASARI (Visually Accessible Rembrandt Images) criteria in the pre-surgical differentiation of IDH-mutant-grade-4 astrocytomas(A4-mut) and IDH-wildtype-glioblastomas(Gb-wt), and the potential added-value of incorporating the T2/FLAIRmismatch into the visual evaluation features.
Methods or Background: Histomolecularly confirmed A4-mut or Gb-wt(age-balanced) with available pre-surgical-MRI were recruited retrospectively (2016-2022) from three tertiary hospitals in Spain and The-Netherlands.
Blinded, independent assessments of VASARI_features and T2/FLAIRmismatch were performed by three-experienced-neuroradiologists (5, 10, and >20years of experience), and a final consensus reading was obtained.
Statistical differences(x2-test) and potential discriminatory performance (AUC-ROC) for each VASARI_features were assessed in univariate analysis.
Additionally, two multivariable logistic-regression models were developed: VASARI-only and VASARI+T2/FLAIRmismatch. The discriminative performance of the models was assessed by AUC-ROC values (accuracy, sensitivity and specificity were derived from Youden-index).
All analyses were validated using 5-fold cross-validation.
Results or Findings: A total of 163 patients were included: 43 A4-mut(mean-age 41y.o.,28males) and 120 Gb-wt(mean-age 47y.o.,87males).
Univariate diagnostic performance: The highest AUC-ROC values (>0.70) were T2/FLAIRmismatch, proportion of non-enhancing tumor, definition of enhancing margin, and thickness of enhancing margin (AUC=0.72-0.76). 30% of variables showed AUC=0.60-0.70 and 50% of variables showed AUC<0.60.
The performance of the multivariable models was as follows: the VASARI-only model (three most discriminative features) achieved an AUC=0.82 (accuracy=0.77); whereas the VASARI+T2/FLAIRmismatch model achieved an AUC=0.87 (accuracy=0.88). Expert neuroradiologists’ consensus accuracy was 0.83.
Conclusion: A4-mut exhibited higher proportions of non-enhancing-tumor, ill-defined enhancing borders, and thinner enhancing margins as the most prominent radiological differences.
The T2/FLAIRmismatch, present in any proportion (in some cases focal and involving <25% of the total MRI abnormality extent), showed almost perfect specificity for A4-mut.
An accuracy of 0.88 was achieved by the multivariable model (VASARIs plus the T2/FLAIRmismatch), exceeding the neuroradiologists’ consensus.
Limitations: None
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: None

Author Block: Y. Liang, Y. Liu, Z. Chen, M. Ge, Y. Wang; Beijing/CN
Purpose: In non-contrast-enhanced (non-CE) adult diffuse gliomas, a significant proportion (19–44%) exhibit aggressive behavior due to high-grade molecular features like IDH wild-type status. Precise preoperative IDH prediction is crucial for optimizing treatment. The 2021 WHO Classification of CNS Tumours designates IDH status as a core determinant for molecular subtyping and grading. IDH wild-type gliomas, even those with low-grade histological appearance, are classified as high-grade and require intensified treatment. Conventional imaging lacks quantitative characterization, whereas radiomics non-invasively reveals tumor heterogeneity by extracting high-throughput quantitative features. This study aims to develop a multimodal MRI radiomics model for preoperative non-invasive prediction of IDH mutation status to guide individualized therapy.
Methods or Background: A retrospective analysis included 151 patients (158 lesions) with pathologically confirmed non-CE adult diffuse gliomas (2016–2023). Based on pathology, patients were stratified into IDH-mutant (44 lesions) and IDH-wild-type (114 lesions) groups and randomly divided into training and validation sets (7:3 ratio). Regions of interest were manually delineated on T2WI, followed by co-registration with ADC and T1CE sequences. A total of 1,132 radiomics features were extracted from T2WI, T1WI, ADC, and T1CE sequences. Feature selection involved ICC (≥0.85), statistical tests, and LASSO regression. Six machine learning models (LR, SVM with RBF/linear kernels, KNN, DT, NB) were evaluated using AUC to assess single-sequence and multi-sequence predictive performance.
Results or Findings: Multi-parametric models combining T2WI, T1CE, and ADC outperformed single-sequence models. The SVM (RBF kernel) classifier with multi-parametric features achieved optimal performance (training AUC = 0.969, sensitivity 88.6%, specificity 100%, accuracy 85.4%; validation AUC = 0.922, sensitivity 82.3%, specificity 90.3%, accuracy 87.3%), significantly surpassing single-sequence models.
Conclusion: The multimodal MRI radiomics model enables accurate, non-invasive IDH status prediction in non-enhancing gliomas (AUC>0.9), assisting surgical planning and advancing imaging-genomic diagnostics.
Limitations: Not applicable
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: I. Jacečková, V. Sedlák, M. Majovsky, A. Kavkova, D. Netuka, T. Belsan, K. Sichova, E. Stoklasa, M. Nemy; Prague/CZ
Purpose: To benchmark the diagnostic performance of conventional and advanced diffusion MRI models, including ADC, DTI, DKI, SMT, NODDI, and RSI, in predicting histologic grade and IDH mutation status in adult-type diffuse gliomas, and to investigate whether model performance differs between various tumor zones and peritumoral regions.
Methods or Background: A cohort of 200 patients with histologically confirmed adult-type gliomas underwent a standardized 3T MRI protocol including multi-shell diffusion imaging. Quantitative parametric maps were reconstructed from six diffusion models (ADC, DTI, DKI, SMT, NODDI, RSI). Tumoral and peritumoral regions were delineated automatically by a convolutional neural network on structural MRI. Zone-specific diffusion parameters were extracted and evaluated for prediction of histologic grade and IDH mutation. ROC analysis and cross-validation were performed to compare model performance, with ADC serving as the clinical benchmark.
Results or Findings: In addition to significantly increasing diagnostic performance in predicting grade and IDH, advanced multi-compartment models revealed marked spatial heterogeneity not captured by conventional ADC. IDH-wildtype gliomas exhibited elevated cellularity indices in both enhancing tumor and peritumoral regions, whereas IDH-mutant gliomas showed more confined abnormalities. Higher-grade gliomas displayed pronounced microstructural disruption extending into edema and peritumoral tissue.
Conclusion: This study establishes a comparative benchmark of diffusion models for glioma characterization. NODDI, SMT, and RSI outperform conventional ADC and show robust performance across both tumoral and peritumoral compartments. By leveraging automated segmentation and multi-shell diffusion analysis, zone-specific biomarkers can be extracted with minimal human input, supporting precision preoperative assessment and guiding model selection in clinical neuro-oncology practice.
Limitations: The main limitations of this study are its single-center nature and still a relatively small sample size, despite being one of the largest advanced diffusion datasets published. Also, less common models (e.g. CHARMED, VERDICT) were not evaluated.
Funding for this study: Funding was provided by the Ministry of Health of the Czech Republic, grant no. NW25J-08-00023, and by Charles University, project
GA UK no. 222623.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Ethics Committee of the Military University Hospital Prague.

Author Block: Q. Zheng, X. Su, Y. Deng, X. Yang, L. Wang, S. Tang, Y. Jin, Q. Yue; Chengdu, Sichuan Province/CN
Purpose: Diffuse midline gliomas, H3 K27-altered (DMGs) are classified as WHO grade 4 tumors regardless of histopathological appearance. Recent advances reveal substantial heterogeneity among these patients, with some exhibiting histological features resembling lower-grade (LGG, histological grade 2-3) group and others resembling high-grade (HGG, histological grade 4) group. This research aims to identify distinct radiomics and genomics features in LGG and HGG groups with H3 K27-altered and link radiomic signatures with relevant genomic alterations.
Methods or Background: A cohort of 104 H3 K27-altered DMG patients, diagnosed between December 2016 and February 2023, were classified as LGG or HGG by experienced pathologists. Of these, 48 patients comprised the radiomics set from which radiomic features were extracted from preoperative MRI and selected using least absolute shrinkage and selection operator (LASSO) regression. A separate radiogenomics set of 27 patients with both MRI and whole-exome sequencing (WES) data was used to explore associations between radiomic and genomic features. The biological meaning of radiomics-associated key genes was explored using a public glioma dataset from the Chinese Glioma Genome Atlas (CGGA).
Results or Findings: Seven radiomic features were finally selected. Among 35 differential mutation genes, original_glszm_LargeAreaHighGrayLevelEmphasis was associated with AATK, GRIN2A, NEFH and TNK2. In CGGA, GSVA analysis revealed enrichment of pathways related to these radiomics-correlated genes’ product. Furthermore, GRIN2A, NEFH and TNK2 emerged as independent prognostic factors for overall survival (OS) in glioma.
Conclusion: Histological grade-specific radiomic features derived from preoperative MRI in H3 K27-altered DMGs are linked to key prognostic genes, offering novel insights into the pathophysiological mechanisms of these tumors and their imaging correlates.
Limitations: This study is limited by its retrospective design and the relatively small sample size. Furthermore, multi-center validation is needed.
Funding for this study: The National Natural Science Foundation of China (Grant No. 82271961) ;
1·3·5 projects for Artificial Intelligence (ZYAI24050), West China Hospital, Sichuan University
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The research protocol was approved by the ethics commissions of the West China Hospital Ethics Committee with a waiver of informed consent (number IRB-2023-745)

Author Block: H. Wu¹, T. Xie²; ¹Chongqing/CN, ²Ch/CN
Purpose: To determine whether hypoperfused yet metabolically active tumor regions—termed low-angiogenic tumor (LAT) habitats—impair glymphatic clearance in the contralateral hemisphere and predict clinical outcomes in patients with high-grade glioma (HGG) undergoing standard chemoradiotherapy.
Methods or Background: Materials and methods: This retrospective study included 151 newly diagnosed HGG patients who underwent preoperative multiparametric MRI and received standardized treatment. LAT habitats were delineated from perfusion maps, and their metabolic activity was estimated using a weighted least squares model based on multi-voxel MRS of enhancing tumor regions. Glymphatic function was assessed via the DTI-ALPS index in the anatomically unaffected hemisphere. Partial correlation, single and serial mediation models, and Cox regression were used to evaluate interdependencies among LAT perfusion, metabolism, glymphatic dysfunction, and progression-free survival (PFS), adjusting for IDH mutation status.
Results or Findings: Results：LAT rCBV, Cho/NAA, and ALPS showed strong intercorrelations (r = 0.763, 0.591, 0.409; all p < 0.001). Mediation revealed a full pathway: LAT perfusion predicted metabolism (β = 1.25, p < 0.001), which predicted glymphatic function (β = 0.06, p < 0.001); only the indirect effect was significant. Cox analysis identified Cho/NAA and IDH—but not ALPS—as independent PFS predictors. Serial mediation confirmed that only metabolism, not glymphatic dysfunction, mediated survival. A multivariable model yielded a C-index of 0.922.
Conclusion: Conclusions: Metabolically active LAT habitats remotely impair glymphatic function and contribute to recurrence. Combined metabolic–glymphatic profiling may offer a mechanistic basis for risk stratification and therapeutic targeting in HGG.
Limitations: LAT metabolism was estimated indirectly, ALPS was measured only contralesionally, and the retrospective design limits causality; moreover, unmeasured solid stress—potentially co-localizing with metabolic tension—may contribute to contralesional glymphatic impairment and warrants prospective multimodal evaluation.
Funding for this study: N/A
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: M. R. Stański, M. Goralewski, J. Watorek, S. Antczak, J. Moskal, K. Katulska; Poznań/PL
Purpose: Low grade gliomas (LGG) are slowly growing tumours which induce neuroplasticity. It was previously shown that the invasion of hippocampus by LGG may induce enlargement of contra-lesional hippocampus, including its gray matter volume (GMV) and volumes of its subfields. Our goal was to assess if similar changes may occur also in patients with LGG which do not invade hippocampus.
Methods or Background: This was a retrospective study of 3D T1 MRI scans of 30 LGG patients (17 left-, 13 right-sided) without hippocampal invasion and 26 healthy controls. In each case a neuroradiologist drew a 3D volume of interest delineating the tumor. We used virtual brain grafting (VBG) to replace abnormal tissue with a clipping of healthy brain template. Using synthetic images we analyzed GMV with voxel-based morphometry (VBM) and volume of hippocampal subfields with FreeSurfer 6.0. We focused on regions previously reported in the literature to be enlarged in LGG with hippocampal infiltration.
Results or Findings: Contra-lesional hippocampal GMV was significantly larger in left- (T = 7.78, p < 0.05) and right-sided LGG (T = 6.39, p < 0.05). A trend towards larger hippocampal-amygdala transitional area (HATA) was seen in both groups, reaching significance in right-sided LGG (Bonferroni corrected p = 0.04). In right-sided LGGs, trends for enlargement were also observed in GC-ML-DG head (the granule cell molecular layer of the dentate gyrus head) , CA1 head, and CA3 head.
Conclusion: The hippocampus contralateral to LGG may undergo structural remodeling even when the ipsilateral one is not affected by tumour infiltration.
Limitations: The study had retrospective design. Our cohort size was small due to epidemiology of LGGs and strict inclusion criteria. Groups were heterogenous regarding tumor volumes due to difficulty in adjusting it with small samples.
Funding for this study: The study received no external funding.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Bioethics Committee of Poznań University of Medical Sciences

Author Block: M. Correia De Verdier¹, R. Saluja², L. Gagnon³, U. Baid⁴, M. Astaraki⁵, R. Huang⁶, S. Bakas⁴, E. Calabrese⁷, J. D. Rudie⁸, B. M. L. G. C. ¹; ¹Uppsala/SE, ²New York, NY/US, ³Québec City, QC/CA, ⁴Indianapolis, IN/US, ⁵Stockholm/SE, ⁶Boston, MA/US, ⁷Durham, NC/US, ⁸La Jolla, CA/US
Purpose: The 2025 Brain Tumor Segmentation (BraTS) challenge on pre- and post-treatment glioma aims to create a large public dataset of annotated diffuse glioma MRIs and a benchmarking environment for developing and evaluating deep learning segmentation models to address challenges in treatment planning and disease monitoring.
Methods or Background: Eighteen institutions on four continents contributed 4401 MRIs of patients with diffuse gliomas, acquired pre- or post-surgery, radiation, or systemic therapy. MRIs included pre-contrast and contrast-enhanced T1-weighted, T2-weighted, and T2-FLAIR sequences. Data preprocessing and annotation followed established BraTS guidelines. Neuroradiologists approved annotations for four sub-regions: enhancing tissue (ET), surrounding non-enhancing FLAIR hyperintensity (SNFH), non-enhancing tumour core (NETC), and resection cavity (RC). Participants used the dataset to develop and evaluate their segmentation models, predicting ET, RC, tumour core (TC = ET + NETC) and whole tumour (WT = TC + SNFH). Forty-four teams participated in the validation phase and 12 teams in the testing phase. Evaluation was performed using lesion-wise Dice Similarity Coefficient (L-DSC) and Normalized Surface Distance (L-NSD).
Results or Findings: The best performing team's L-DSC (mean (median) ± SD) were: ET – 0.81 (0.93) ± 0.26, RC – 0.89 (1.00) ± 0.25, TC – 0.81 (0.94) ± 0.28, WT – 0.88 (0.96) ± 0.18, and L-NSD were: ET – 0.86 (0.96) ± 0.25, RC – 0.89 (1.00) ± 0.25, TC – 0.82 (0.94) ± 0.27 and WT – 0.85 (0.94) ± 0.19.
Conclusion: The 2025 BraTS challenge sets a benchmark for deep learning segmentation models using the largest expert-annotated glioma dataset. The winning team achieved excellent performance with high L-DSC and L-NSD. The developed models may aid objective tumour assessment thought the patient’s entire clinical course.
Limitations: The annotator model does not directly take into account inter-observer variability.
Funding for this study: Not applicable.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Institutional review board approval from each participating institution.

Author Block: E. Zhao¹, X. Liu², Y. Shi¹, X. Gao¹, X. Zheng¹, C. Yang³, J. Liu¹; ¹TianJin/CN, ²Tianjin/CN, ³Dalian/CN
Purpose: To develop a deep learning–driven model integrating intratumoral and peritumoral features for survival prediction in glioblastoma. Risk stratification was performed by combining a deep learning score (DL score) with an intratumoral heterogeneity score (ITH score) derived from habitat analysis, and their biological underpinnings were explored.
Methods or Background: We retrospectively included 511 pathologically confirmed glioblastoma patients from three hospitals (n=381, model development) and the TCGA cohort (n=130, biological validation). Tumor, peritumoral 10 mm, 20 mm, and edema regions were segmented on contrast-enhanced T1WI and T2WI. Each region was clustered into three habitats using k-means. Multiscale habitat features were quantified to calculate ITH scores. A 3D ResNet101 with Cox partial likelihood loss (DeepSurv) extracted global features to predict survival probabilities. Model performance was evaluated by concordance index (C-index), and DL scores were derived from prediction probability. Patients were dichotomized into high- and low-risk groups by median cutoff, and Kaplan–Meier survival analysis was performed across all DL and ITH scores. Groups with significant stratification (log-rank p<0.001) underwent biological validation in TCGA, including whole-exome sequencing,copy number alterations, tumor mutation burden, RNA-seq, methylation, and proteomics.
Results or Findings: The combined tumor and 10 mm peritumoral model achieved robust prognostic accuracy (validation C-index=0.75; 1- and 2-year AUCs >0.85). Both intratumoral DL score and intratumoral/peritumoral_10mm ITH scores independently stratified survival (p<0.001). High DL scores were linked to oncogenic pathway activation (MAPK, VEGF, focal adhesion) and elevated tumor–stroma ratio, while high ITH scores reflected immune heterogeneity, low tumor mutation burden, and enriched checkpoint expression. Joint stratification (high DL + high ITH) identified the poorest subgroup.
Conclusion: Integrating global deep learning features with intratumoral and peritumoral heterogeneity provides a novel, biologically interpretable basis for risk stratification in glioblastoma prognosis.
Limitations: Validation is biologically indirect, not direct
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Tianjin Medical University General Hospital (IRB2025-YX-446-01)
The Second Affiliated Hospital of Dalian Medical University (KY2024-029-01)
The First Affiliated Hospital of Dalian Medical University (PJ-KS-KY-2023-422)

Author Block: M. Choi¹, Y. Choi², J. Lee¹, M. Kim¹, I. Hwang¹, Y. W. Park¹, J. E. Park¹, S. H. Choi¹, K. Choi¹; ¹Seoul/KR, ²Pocheon/KR
Purpose: The 2021 WHO classification redefines Grade 2/3 IDH-wildtype diffuse astrocytic tumors with specific molecular alterations as molecular glioblastoma (mol-GBM), grade 4. Since mol-GBM lacks the histological hallmarks of conventional glioblastoma (hist-GBM), it is hypothesized to exhibit a distinct pattern of tumor heterogeneity. This study aimed to explore how mol-GBM differs from hist-GBM in tumor heterogeneity using multi-parametric physiologic MRI and provide novel radiologic insights into the characteristic features of this newly defined subtype.
Methods or Background: In this multi-institutional retrospective study, imaging data were collected from two tertiary centers: 13 mol-GBMs and 39 hist-GBMs from institution 1 (2007–2024) for development, and 9 mol-GBMs from institution 2 (2020–2024) for external validation. Apparent diffusion coefficient (ADC; cellularity), relative cerebral blood volume (rCBV; vascularity), and volume transfer constant (Ktrans; permeability) were binarized into high/low categories, yielding eight spatial habitat clusters (2×2×2) to visualize tumor heterogeneity. Habitat distributions and histogram features (mean, median, 10th, 90th percentile) were compared between groups. Diagnostic performance was assessed by multivariable logistic regression with ROC analysis.
Results or Findings: The tumor volume of high Ktrans clusters was significantly smaller in mol-GBM than hist-GBM (median; 2.5cm³ vs. 14.6cm³, p<0.001), particularly for the most malignant cluster (low ADC, high rCBV, high Ktrans; 0.06cm³ vs. 7.0cm³, p<0.001). Mol-GBM showed lower mean and 90th percentile values of rCBV (p=0.013) and Ktrans (p<0.001), whereas no difference was observed for ADC. Consequently, adding rCBV and Ktrans to ADC significantly improved differentiation, increasing the AUC from 0.70 to 0.89 (p=0.002).
Conclusion: Non-invasive characterization of tumor heterogeneity using multi-parametric MRI habitat analysis revealed distinct patterns in mol-GBM versus hist-GBM. The most discriminating parameter was tumor permeability (Ktrans), with mol-GBM showing smaller hyperpermeable habitats. In contrast, cellularity (ADC) was comparable across both subtypes.
Limitations: Not applicable.
Funding for this study: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00251022) (K.S.C.); the Phase III (Postdoctoral fellowship) grant of the SPST (SNU-SNUH Physician Scientist Training) Program (K.S.C.); the SNUH Research Fund (No. 04-2024-0600) (K.S.C.); and the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) grant funded by the Ministry of Health&Welfare (No. RS-2024-00439549) (K.S.C.).
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: C. García¹, I. MARTINEZ-ZALACAIN², P. Naval-Baudin², A. Camins Simó², A. Jareño Badenos², M. Cos Domingo², C. Majós², A. Pons Escoda²; ¹Murcia/ES, ²Hospitalet de Llobregat/ES
Purpose: To compare the diagnostic performance of CBV- and PSR-optimized sequences in the presurgical differentiation of glioblastoma, brain-metastasis, lymphoma, and meningioma, and to assess the additive value of combining both metrics within and across the sequences.
Methods or Background: Retrospective single-center cohort: glioblastoma (n=121), metastasis (n=63), meningioma (n=55), lymphoma (n=13). Each patient underwent two consecutive DSC-acquisitions within the same MR-session: 1st- non-preloaded, high flip-angle (PSR-optimized); and 2nd- full-dose preloaded (using prior contrast), intermediate flip-angle (CBV-optimized, standardized, consensus-compliant). From enhancing tumor masks (plus edema masks for glioblastoma vs. metastasis), we extracted rCBV (NAWM-normalized, leakage-corrected) and PSR statistics. For each tumor-pair we identified the best single-acquisition metric, and trained bivariate logistic models to combine CBV and PSR both within and across both protocols.
Results or Findings: Single-metric AUCs for the 1st DSC-sequence ranged 0.72 (Gb_vs_Metastasis, nrCBVmin_edema)- 0.91 (Meningioma_vs_Lymphoma, PSRp75), average AUC=0.79. For the 2nd, AUCs ranged 0.72 (Gb_vs_Metastasis, PSRmax_edema)- 0.89 (Meningioma_vs_Lymphoma, nrCBV_p75), average AUC=0.80. Overall, CBV and PSR were the best metrics in 50% of comparisons each. Selecting the best-performing protocol per tumor-pair increased average AUC to 0.82. The bivariate intraprotocol models combining nrCBV+PSR in the 1st DSC AUCs ranged 0.76 (Gb_vs_Metastasis, PSRmean_edema+nrCBVmin_edema)- 0.94 (Meningioma_vs_Lymphoma, PSRp75+nrCBVp75), average AUC=0.85. In the 2nd DSC AUCs ranged 0.76 (Gb_vs_Metastasis, PSRmax_edema+nrCBVmax_edema)- 0.93 (Meningioma_vs_Lymphoma, PSRp75+nrCBVp75), average AUC=0.83 The bivariate cross-protocols models combining nrCBV+PSR AUCs ranged 0.8 (Gb_vs_Metastasis, PSRmin_enhancing_2nd+PSR_mean_edema_1st) to 0.94 (Meningioma_vs_Lymphoma, PSRp75_2nd+nrCBVp75_1st), average AUC=0.87.
Conclusion: Both sequences performed similarly in pairwise tumor classification, with CBV and PSR showing no clear predominance in either acquisition. Within each protocol, models combining CBV and PSR outperformed single-metric approaches, with optimal performance achieved when integrating both protocols. Our findings indicate that implementing the dual-DSC protocol in clinical practice could maximize the accuracy of presurgical diagnosis.
Limitations: Single-centre, retrospective design.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Research Ethics Committee of Hospital Universitari de Bellvitge.

Author Block: S. Suvak, B. Papazov, E. Schulz, J. Ricke, F. Ringel, S. Stöcklein, V. Stöcklein; Munich/DE
Purpose: Accurate determination of language lateralization is critical for surgical planning in glioma patients, particularly when tumors involve the dominant hemisphere. Task-based fMRI is currently used but is limited by availability and patient compliance. Resting-state fMRI (rs-fMRI) may provide a feasible alternative for identifying the language-dominant hemisphere.
Methods or Background: 22 glioma patients prospectively underwent structural MRI (3D T1w and FLAIR), task-based fMRI (sentence generation), and rs-fMRI (6 min each). Broca’s and Wernicke’s areas and tumors were semi-automatically segmented. For task-fMRI, activation volumes were determined, while rs-fMRI assessed volumes of regions functionally connected to bilateral Broca seeds (r = 0.35). Language lateralization was quantified using the lateralization index (LI = [L–R]/[L+R]), with values near ±1 indicating strong lateralization. Task- vs. rs-fMRI-derived LI were compared using paired t-tests and Pearson correlation.
Results or Findings: 22 patients (56 ± 18 yrs) were analyzed; two were excluded due to motion artifacts. Hemispheric lateralization was concordant between task-based and rs-fMRI in 19 patients (95.0%). Rs-fMRI identified left-hemispheric dominance in 17 (mean LI = 0.28 ± 0.17) and right-hemispheric dominance in 3 (LI = –0.11 ± 0.15), while task-fMRI showed left dominance in 18 (LI = 0.49 ± 0.22) and right dominance in 2 (LI = –0.42 ± 0.41). A strong positive correlation was observed between methods (r = 0.69, p = 0.0007).
Conclusion: Rs-fMRI identified the language-dominant hemisphere in 95% of glioma patients, showing strong concordance with task-based fMRI. As a non-task-dependent, brief, and automatable technique, rs-fMRI may serve as a standardized alternative for preoperative language mapping, facilitating surgical planning and risk assessment.
Limitations: This is a single-center study and intraoperative mapping will be needed to further confirm language lateralization.
Funding for this study: This work is the result of a research cooperation with Brainlab SE, Munich, Germany. Otherwise, all authors declare no conflict of interest regarding the materials used or the results presented in this study. All authors declare no other relationships or activities that could appear to have influenced the submitted work.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: We performed the study in accordance with the Declaration of Helsinki and the STROBE statement. All patients included in this study provided written informed consent.

Author Block: Y. Li, X. Hong, J. Li, Z. Zhuo, R. Zhang, C. Ye, Y. Liu; Beijing/CN
Purpose: Preoperative neuropathological evaluation of adult-type diffuse gliomas (ADG) is crucial for guiding diagnosis and treatment. We aimed to develop an MRI-based glioma neuropathology prediction (MRI-GNP) deep learning system and assess the ability of deep learning methods to comprehensively predict neuropathology markers of ADGs.
Methods or Background: We utilized 35,616 MR images of 8,844 patients across 22 datasets, along with 39,642 corresponding neuropathology markers. We evaluated various deep learning architectures, input formats, and training strategies to identify the optimal configuration. Model performance was assessed using accuracy and area under the curve (AUC). We evaluated the model’s ability to predict 12 tasks. Furthermore, we evaluated MRI-GNP to improve the diagnostic accuracy of neuroradiologists. Finally, to address the challenge of missing contrast-enhanced T1-weighted imaging (T1CE) sequences, we integrated MRI-GNP with generative models.
Results or Findings: Pretrained vision transformers with a 2.5D input configuration were selected. The model achieved high performance (AUC ≥ 0.8) on several prediction tasks, including WHO grade 4 (AUC = 0.852), Ki-67 expression (AUC = 0.817), IDH mutation (AUC = 0.826), and 1p/19q codeletion (AUC = 0.823) based on test sets, whereas for other tasks, such as WHO grade 2/3/4, +7/-10 alteration, CDKN2A/B homozygous deletion, TERT promoter mutation, and EGFR amplification, the model demonstrated moderate performance (AUC ≥ 0.7). The model's performance was poor for MGMT promoter methylation, TP53 mutation, and ATRX mutation. Furthermore, MRI-GNP was validated to significantly improve diagnostic accuracy of neuroradiologists. Finally, incorporating synthetic images enabled MRI-GNP to maintain comparable performance in the absence of T1CE sequences.
Conclusion: MRI-GNP is a robust and generalizable deep learning system for preoperative neuropathology prediction, which has strong potential to enhance precision diagnostics and support clinical decision-making.
Limitations: No limitations were identified.
Funding for this study: Funding was provided by Beijing Excellent Young Scientists Program in Higher Education, Beijing Hospital Management Center-Climb Plan, General Program of Beijing Natural Science Foundation, Youth Program of Beijing Natural Science Foundation, Capital Health Development Scientific Research Special Project of Beijing Municipal, and Radiation Imaging Database Project of the National Health Commission.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Institutional Review Board of Beijing Tiantan Hospital, Capital Medical University, Beijing, China (No. 82202084, QML20210505 and 82330057).