Non-Invasive Differentiation of Diffuse Midline Glioma and Midline Glioblastoma Using DCE-MRI Perfusion Parameters and Machine Learning Classification in Pediatric and Young Adult Patients

Non-Invasive Differentiation of Diffuse Midline Glioma and Midline Glioblastoma Using DCE-MRI Perfusion Parameters and Machine Learning Classification in Pediatric and Young Adult Patients | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Non-Invasive Differentiation of Diffuse Midline Glioma and Midline Glioblastoma Using DCE-MRI Perfusion Parameters and Machine Learning Classification in Pediatric and Young Adult Patients Anshika Kesari, Rakesh Kumar Gupta, Sunita Ahlawat, Rana Patir, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8481592/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Purpose Diffuse midline glioma (DMG) and midline glioblastoma (mGBM) are aggressive WHO grade 4 tumors with comparable median survival of 12–18 months but require fundamentally different therapeutic approaches. Despite their clinical urgency, non-invasive differentiation remains challenging due to overlapping conventional MRI features and the difficulty of obtaining tissue diagnosis from eloquent midline locations. Methods This retrospective study included 62 patients with histologically confirmed midline gliomas (30 mGBM, 32 DMG) evaluated with 3T MRI. Quantitative DCE-MRI perfusion parameters (rCBV, rCBF, Slope-2, K trans , V p , V e ) were computed and compared between the midline tumor types. Statistical analyses included Shapiro-Wilk test, t-test, and ROC curve analysis using perfusion parameters. Machine learning-based classification was also performed using four classifiers and 5-fold cross-validation, evaluating all possible feature combinations among the best features from the perfusion parameters. Results The 95th percentile values of perfusion parameters demonstrated superior discriminative capability between mGBM and DMG. DMG exhibited significantly lower perfusion parameter values compared to mGBM (p < 0.05). Individual perfusion parameters, particularly rCBF, rCBV, V e showed discriminative performance achieving AUC values ranging from 70.62% to 75.31%, for differentiating mGBM vs DMG. Machine learning classifiers used these features for evaluating 7 combinations. Three parameter combination (rCBV + rCBF + V e ) using RF achieved highest cross-validation accuracy (76.67 ± 7.16%) with consistent sensitivity (80.00%) across all models. Conclusions Quantitative DCE-MRI perfusion analysis provides significant diagnostic value for differentiating DMG from mGBM, offering a non-invasive alternative when tissue diagnosis is not obtained. Both individual parameters and optimized multi-parametric approaches demonstrate clinically useful performance for guiding treatment decisions. Midline glioblastoma Diffuse midline glioma Dynamic contrast-enhanced MRI Perfusion parameters Machine learning Sequential forward selection Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1 Introduction H3 K27M-mutant diffuse midline glioma (DMG) was recognized as a distinct brain tumor type in the 2016 World Health Organization (WHO) Classification of Tumors of the Central Nervous System. This classification was updated in 2021, renaming the entity as H3 K27-altered DMG defined by mutations affecting the histone H3 protein resulting in K27M amino acid substitution [1,2]. This molecular definition supersedes histological grading, with all H3 K27-altered DMGs considered WHO grade 4, irrespective of histological features or location. Although DMGs have been classified as pediatric-type diffuse high-grade gliomas (HGGs), they are also frequently diagnosed in young adults and occasionally in older patients [3]. These tumors primarily arise in central midline structures including the thalamus, brainstem, and spinal cord [3] and are associated with poor prognosis [4–6]. However, accurate diagnosis of DMG can be challenging due to imaging similarities with other aggressive midline tumors, particularly midline glioblastoma (mGBM), despite their distinct molecular profiles and treatment requirements. Accurate differentiation between DMG and mGBM has critical therapeutic and prognostic implications despite their similarly aggressive nature. Both entities are WHO grade 4 tumors with comparable median survival of approximately 12-18 months [2,4,7]. However, they require fundamentally different treatment strategies with significant implications for patient management and clinical outcomes. Standard treatment for mGBM consists of maximal safe resection followed by concurrent temozolomide chemotherapy and radiotherapy (Stupp protocol), with adjuvant temozolomide maintenance [8–10]. This approach demonstrates survival benefit in appropriately selected patients. In contrast, DMG management emphasizes radiation therapy as the primary treatment modality, as extensive surgical resection is often precluded by eloquent tumor locations [5,8,9]. Importantly, emerging targeted therapies including histone deacetylase (HDAC) inhibitors and ONC201 show specific promise for H3 K27-altered DMG, making accurate preoperative identification essential for clinical trial eligibility and personalized treatment planning [5,10,11]. Given these therapeutic implications, reliable preoperative distinction between these entities is essential. The gold standard for definitive differentiation relies on molecular testing for H3 K27-alteration through immunohistochemistry or genetic sequencing of tumor tissue [2,4]. However, obtaining adequate tissue for analysis presents substantial clinical challenges due to the eloquent midline locations of these tumors [9,12–14]. Stereotactic biopsy in these regions carries considerable risks of neurological complications, hemorrhage, and procedure-related morbidity, and in many cases, surgical intervention is contraindicated due to critical anatomical location or patient frailty [9,12,13]. Furthermore, conventional magnetic resonance imaging (MRI) sequences, including T1-weighted (T1-W), T2-weighted (T2-W), and fluid-attenuated recovery (FLAIR) images, demonstrate significant limitations in distinguishing these entities. Both tumor types characteristically present as heterogeneously enhancing masses with surrounding edema and variable necrosis on conventional imaging [9,15–17]. While strict midline location may suggest DMG, substantial overlap in imaging appearances between mGBM and DMG complicates reliable non-invasive differentiation [12–14,17,18], underscoring the need for advanced imaging biomarkers. To address this clinical need, advanced MRI techniques beyond conventional sequences have emerged as promising non-invasive tools. MR diffusion-weighted imaging (DWI) provides cellularity information through apparent diffusion coefficient (ADC) mapping [18–25]. MR perfusion-weighted imaging (PWI), including both dynamic susceptibility contrast (DSC) [12,19,24,26] and dynamic contrast-enhanced (DCE) MRI [27–29], offers complementary information about tumor vascularity. Among these, DCE-MRI enables quantitative assessment of tissue vascularity and permeability by tracking the kinetics of contrast agent distribution. Quantitative perfusion parameters derived from DCE-MRI, including relative cerebral blood volume (rCBV), and relative cerebral blood flow (rCBF), K trans (volume transfer constant), K ep (rate constant), V e (extravascular extracellular space volume fraction), and V p (plasma volume fraction), provide insights into tumor angiogenesis and blood-brain barrier integrity [18,24,30–33]. Previous studies have examined various imaging approaches for DMG characterization, ranging from conventional MRI features [16,17,34] to advanced techniques including diffusion [18,20,22,24] and perfusion imaging [12,24,26]. Recently, classical machine learning (ML) methods utilizing perfusion features have been applied for tumor classification, showing that combining multiple DCE-MRI parameters can improve classification accuracy compared to single parameter approaches [14,20,35]. However, comprehensive studies directly comparing individual DCE-MRI quantitative perfusion parameters and systematic evaluation of multi-parametric ML classifiers specifically for distinguishing DMG from mGBM in children and young adults remain limited. Addressing these gaps, the present study aims to evaluate the utility of quantitative DCE-MRI perfusion parameters using both individual parameters and multi-parametric MRI-based machine learning classifiers for distinguishing DMG from mGBM in children and young adults. These entities represent a particular diagnostic challenge due to their overlapping anatomical distribution and similar conventional imaging features, including contrast enhancement patterns and the presence of necrosis. We systematically assess the discriminatory performance of individual perfusion metrics and optimized ML combinations, and their relationship to molecular features, particularly H3 K27-alteration status. 2 Materials and Methods 2.1 Patient Population The dataset was acquired from the XYZ hospital. The retrospective study protocol was approved by the institutional ethics committee, and the requirement for informed consent was waived by the committee. Patients with HGG and MRI data between 2018-2024 were identified. The study applied strict inclusion criteria: (1) histo-pathologically confirmed diagnosis of DMG, or mGBM; (2) availability of preoperative DCE-MRI data along with conventional MRI; and (3) molecular testing results for IDH mutation or H3 K27-alteration (Figure 1). Patients were excluded beforehand if they had prior treatment, significant motion artefacts on MRI, incomplete clinical data or any other type of tumor. In the current study, 62 patients satisfied the patient selection criteria including age (< 40 years) criteria. Figure 1 shows the final cohort that met the inclusion and exclusion criteria (30 mGBMs, and 32 DMGs). Patient demographics, tumor location, and molecular characteristics were extracted from their records. Tumors were classified according to WHO Classification of Central Nervous System Tumors [1,2]. 2.2 MRI Acquisition Protocol All imaging studies were performed on a 3T whole-body MRI system (Ingenia, Philips Healthcare, Netherlands) using a 15-channel head coil. The MRI protocol included axial T1-W turbo-spin echo (TSE), axial dual PD-T2-W (TSE), axial FLAIR followed by 3D DCE-MRI (T1-W perfusion MRI) using fast-field echo (FFE) and axial post contrast T1-W (T1-Gd).The contrast injection was administered after the acquisition of four baseline scans to establish pre-contrast signal intensity. DCE-MRI acquisition included 32 dynamic acquisitions with a temporal resolution of 3.8 seconds. A gadolinium-based contrast agent Dotarem (Gadoterate Meglumine, Guerbet, France) was administered intravenously at a dose of 0.1 mmol/kg body weight at a rate of 2.5 ml/sec using a power injector, followed by a 20 ml saline flush. Acquisition parameters for each sequence are listed in Table 1. 2.3 Histopathological and Conventional MRI Analysis The conventional MRI features were assessed by an experienced neuroradiologist with 38 years of experience in neuroimaging, along with histopathological and molecular results. The following features were assessed: (1) tumor location; (2) patient age; (3) H3K27-alteration status; (4) presence of contrast enhancement; and (5) necrosis. Table 1. Detailed protocol for MRI sequences used in the current study. MRI Sequences TR/TE (ms) Flip Angle (ϴ) No. of slices Slice Thickness (mm) FOV (mm 2 ) Acquisition Matrix T1-W 360/10 90 20 6 230 × 230 244 x 237 T2-W 3500/90 90 20 6 230 × 230 244 x 237 PD-W 3500/23.2 90 20 6 230 × 230 244 x 237 FLAIR 4800/340 90 25 0.9 250 × 250 224 x 223 T1-Gd 700/35 80 25 0.9 250 × 250 280 x 278 DCE-MRI 6.38/3.09 20 20 6 230 × 230 192 x 177 Abbreviations: MRI, magnetic resonance imaging; TR, repetition time; TE, echo time; FOV, field of view; T1-W, T1-weighted; T2-W, T2-weighted; PD-W, proton density-weighted; FLAIR, fluid-attenuated inversion recovery; T1-Gd, postcontrast T1-W; DCE-MRI, dynamic contrast-enhanced magnetic resonance imaging. 2.4 DCE-MRI Data Processing The comprehensive pipeline of the whole process is shown in Figure 2. DCE-MRI data processing was performed for quantitative analysis using an in-house developed MATLAB-based software solution (R2021a, MathWorks Inc., Natick, MA, USA) [36,37]. The processing pipeline included the following steps: (1) 3D rigid body registration of all images with respect to T1-W images using SPM12 (Statistical Parametric Mapping software) [38]; (2) De-sculping, and segmentation of gray matter, white matter, and cerebrospinal fluid using SPM12 followed by spatial smoothing; (3) Intensity scaling of structural MRI images (T1-W, T2-W and PD-W) to obtain true pixel signal intensity. (4) Computing pre-contrast T1 maps using T1-W and dual PD-T2-W. (5) Conversion of signal intensity time curves to concentration time curves. (6) Local arterial input function was estimated using the previously reported algorithm for reducing the partial volume effect [36]. (7) Concentration time curves were analyzed voxel-wise to compute various quantitative perfusion parameters using different mathematical models. For the quantification of perfusion parameters, first-pass analysis was performed using the first pass of the contrast bolus through the brain vasculature to compute hemodynamic parameters, including CBV and CBF [37]. Piecewise linear model fitting of the complete concentration time curve was used for the estimation of bolus arrival time (BAT), time to peak during first pass of bolus (BETA), slopes of second (wash-in slope; Slope1) and third line segments (wash-out slope; Slope-2) [36]. The extended Tofts model was applied to estimate the pharmacokinetic parameters, including K trans , V e , and V p using the bidirectional transport of contrast agent between the intravascular and extravascular extracellular spaces [28,36]. Hemodynamic parameters were normalized with respect to the average values of these parameters in contralateral normal-appearing white matter region to obtain their relative values (rCBV, rCBF), which compensates for potential variations in contrast administration and physiological factors across patients [39]. 2.5 Tumor Segmentation Tumor segmentation was performed on FLAIR images using an artificial intelligence (AI) algorithm based on a U-Net architecture [40]. The initial automated segmentation was subsequently reviewed and, when necessary, manually refined by an experienced neuroradiologist. The resulting tumor masks encompassed the entire abnormal signal intensity on FLAIR images, including necrosis, enhancing and non-enhancing components, as well as peritumoral edema. 2.6 Large Blood Vessel (LBV) Segmentation For the segmentation of non-tumoral LBVs, we employed an automated approach based on the combined analysis of rCBV and Slope-2 maps derived from DCE-MRI [41]. The resulting LBV mask was then subtracted from the tumor mask to generate a refined tumor mask excluding non-tumoral vasculature. The refined tumor mask (after LBV removal) was subsequently applied to all parametric maps for further statistical analysis. 2.7 Statistical Feature Selection Statistical features were extracted from the tumor regions on each parametric map (rCBV, rCBF, Slope-2, K trans , V e , and V p ) after LBV removal. These features included minimum value, maximum value, mean, standard deviation, 10 th , 75 th and 95 th percentile. Thus, for further analysis, a total of 42 features (6 parameters × 7 statistical measures) were initially computed. Moreover, among all the slices containing tumor regions, the study was performed on 5 consecutive slices around the center of the tumor to reduce bias in the overall dataset. 2.8 Statistical Analysis The normality of feature distributions within each glioma subtype was assessed using the Shapiro-Wilk test. For each feature, data were grouped according to class, and if all groups passed the test for normality, t-test was performed. If normality was not established in any group, the non-parametric Mann Whitney U-test was used for comparing distributions of perfusion parameters between mGBM and DMG groups. Features with p-value less than 0.05 in the respective test were considered statistically significant. Receiver operating characteristic (ROC) curve analysis was also performed to evaluate the discriminatory performance of each individual perfusion parameter in differentiating between mGBM and DMG. For each parameter, area under the ROC curve (AUC), sensitivity, specificity, and optimal cutoff values were calculated based on Youden's index, and 95% confidence intervals using the DeLong method. Statistical significance was assessed at p < 0.05. Box-and-whisker plots and violin plots were generated to visualize the distribution of 95 th percentile values across tumor types, displaying median, mean, quartiles, and outliers. 2.9 Machine Learning Based Classification To evaluate multi-parametric approaches for glioma subtype differentiation, a comprehensive supervised machine learning (ML) framework was implemented using multiple classical classifiers with rigorous feature selection. 2.9.1 Data Preparation and Splitting The dataset was split into training (70%) and testing (30%) subsets, stratified by class labels to preserve class distribution. Thus, in this study, among 62 patient’s dataset, 43 (21 mGBM, 22 DMG) were used for training and 19 (9 mGBM, 10 DMG) for testing. Prior to model training, standardization was performed via z-score scaling to ensure equal feature importance. 2.9.2 Feature Selection Strategy To identify optimal feature subsets, we systematically evaluated all possible combinations of the top three selected 95 th percentile perfusion parameters (rCBV, rCBF, V e ). This exhaustive approach comprehensively tested single (3 combinations), two (3 combinations), three (1 combination) feature models. This resulted in 7 unique feature combinations evaluated across all four ML classifiers, yielding 28 total model configurations (7 combinations × 4 classifiers). 2.9.3 Classification Models Four ML algorithms were implemented for binary classification. Random forest (RF) with 100 estimators, logistic regression (LR) with L2 regularization and 500 maximum iterations using L-BFGS (Limited-memory Broyden–Fletcher–Goldfarb–Shanno) solver, support vector machine (SVM) using radial basis function kernel with probability estimates enabled, and k-nearest neighbors (KNN) with 5 nearest neighbors, distance weights, and manhattan metrics. All models used random state 42 for reproducibility. Prior to final evaluation, hyperparameter optimization was performed using GridSearchCV with 5-fold cross-validation (CV) to identify optimal model configurations. 2.9.4 Model Evaluation Model performance was comprehensively assessed using 5-fold stratified CV on the training set. The best feature subset for each model-class pair combination was selected based on maximum CV accuracy. Final model performance was evaluated on the held-out validation set using accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). Statistical significance was assessed at p < 0.05. 2.10 Model Implementation All statistical and ML-based analyses were performed in Python (v3.9.12, 64-bit, Anaconda distribution) on a system with an AMD Ryzen 7 5700U processor with Radeon Graphics and 16 GB of memory. Data preprocessing was carried out using pandas (v1.4.2) and NumPy (v1.21.5). Machine learning workflows were implemented with scikit-learn (v1.0.2), while statistical analyses used SciPy (v1.7.3) and stats models (v0.13.2). Visualizations were generated with Matplotlib (v3.5.1) and Seaborn (v0.11.2). 3 Results 3.1 Patient Characteristics The demographic and clinical characteristics of the study population are summarized in Table 2. DMG patients were significantly younger (mean age: 23.9 ± 15.1 years) compared to mGBM (30.5 ± 9.3 years) patients (p = 0.239). Gender distribution showed male predominance in mGBM (63.3% male) with a more balanced distribution in DMG (56.3% male). Molecular profiling showed all DMGs (n=32) were H3 K27-altered while all mGBMs (n=30) cases were lacking this alteration (p < 0.001). Both groups were exclusively IDH-wildtype, consistent with their classification as WHO grade 4 tumors. Tumor location varied significantly between groups, with DMGs predominantly involving midline structure including thalamus (n=18, 56.3%), and brainstem (n=14, 43.7%). while mGBMs were distributed across supratentorial midline locations including thalamus (n=19, 63.3%), corpus callosum (n=7, 23.3%), and other midline structures (n=4, 13.3%) (p < 0.001). Table 2. Demographic, clinical, and conventional MRI characteristics of the study population. Characteristics mGBM DMG Location Supratentorial midline structure Midline structure Age (Mean ± SD years) 30.5 ± 9.3 23.9 ± 15.1 Gender (Males/Females) 19/11 18/14 IDH status Wildtype Wildtype H3K27-alteration Absent Present Contrast enhancement 29/30 23/32 Necrosis 29/30 11/32 3.2 Conventional MRI Features The conventional MRI characteristics showed significant differences between tumor types (Table 2). Contrast enhancement was observed predominantly in mGBM (96.7%) but present in only 71.8% of DMG cases (p < 0.001), with necrosis also most frequent in mGBM (p < 0.001). Figure 3 presents representative examples of mGBM and DMG cases, illustrating both enhanced variants (Case 1) showing strong gadolinium uptake, and non-enhanced variants (Case 2) demonstrating minimal or absent enhancement despite their aggressive WHO grade 4 classification. The figure displays each variant’s FLAIR, T1-Gd, rCBV, and rCBV maps after removing the LBVs. This visualization highlights the challenges of differentiating between mGBM and DMG using qualitative assessment of conventional imaging alone, emphasizing the need for quantitative perfusion analysis. 3.4 Statistical-based Classification Among all statistical features extracted from parametric maps, the 95 th percentile values demonstrated the highest discriminatory power for differentiating between mGBM and DMG. In Figure 4, both the box and whisker plots (Figure 4a) and violin plots (Figure 4b) demonstrate clear separation between the two midline tumor types across the 95 th percentile values of rCBV, rCBF, and V e . Both plots reveal a significant reduction in the median, mean, and their standard deviation values across DMG from mGBM. For rCBV, mGBM exhibited significantly higher values (mean 95 th percentile: 2.11 ± 1.15) compared to DMG (1.45 ± 0.39). Similarly, rCBF and V e values showed a statistically significant decrease in DMG from mGBM with minimal overlap between groups, but the medians remain distinctly separated. The violin plots further confirmed that the distribution density for mGBM is skewed toward higher values, whereas DMG exhibits a more compact distribution centered at lower values. Normality of the data distributions for each glioma subtype was assessed using the Shapiro-Wilk test. In this study, the test results indicated that the assumption of normality (p < 0.05) across features and classes was right and given these findings, parametric statistical methods were employed. Specifically, the t-test was applied which confirmed statistically significant differences for all perfusion parameters between mGBM and DMG (p < 0.05). These significant group differences are summarized in Table 3. ROC curve analysis was performed to quantitatively assess the discriminatory performance of each parameter in differentiating between mGBM and DMG. Figure 5 and Table 3 present the discriminatory performance of individual perfusion parameters for distinguishing the midline HGG subtypes. Figure 5a displays ROC curves for all six perfusion parameters and Figure 5b presents confusion matrices for the three best-performing parameters (rCBV, rCBF, and V e ). Table 3. Statistical-based analysis performance metrics of individual perfusion parameters for distinguishing mGBM from DMG. Perfusion Parameters AUC (%) t-test Sensitivity (%) Specificity (%) 95% CI Optimal Threshold rCBF 75.31 p=0.003** 78.12 73.33 0.631-0.875 29.777 rCBV 71.25 p<0.001*** 78.12 63.33 0.583-0.842 1.621 V e 70.62 p=0.008** 65.82 80.00 0.576-0.836 0.040 Slope 2 65.16 p=0.013* 84.38 60.00 0.514-0.789 0.014 V p 65.00 p<0.013* 87.50 53.33 0.513-0.787 0.018 K trans 55.83 p=0.186 34.38 90.00 0.414-0.702 0.037 Table 3 summarizes the classification metrics for each individual perfusion parameter, such as, rCBF emerged as the best single discriminator, achieving an AUC of 75.31% (95% CI: 0.631-0.875, p<0.001) with sensitivity of 78.12% and specificity of 73.33% at an optimal threshold of 29.777. However, K trans demonstrated the lowest discriminatory power with AUC of 55.83% (p=0.434), indicating it does not significantly distinguish between mGBM and DMG at conventional significance levels. 3.5 Machine Learning-based Classification To investigate the potential of combining multiple quantitative perfusion parameters, we systematically evaluated all possible feature combinations of the statistically performing top three perfusion parameters (rCBV, rCBF, V e ) using four ML classifiers (RF, LR, SVM, KNN). Table 4 presents comprehensive results from this systematic analysis. Table 4. Machine learning based classification performance metrics for distinguishing mGBM from DMG. Selected Features Models CV Accuracy (%) Test Accuracy (%) Precision (%) Recall (%) F1 (%) AUC (%) rCBV RF 69.72 ± 5.72 68.42 66.67 80.00 72.73 74.44 rCBF SVM 72.22 ± 9.17 68.42 66.67 80.00 72.73 73.33 V e RF 74.44 ± 8.90 68.42 66.67 80.00 72.73 58.89 rCBV + rCBF KNN 69.44 ± 10.54 73.68 72.73 80.00 76.19 72.78 rCBV + V e LR 74.72 ± 7.78 57.89 57.14 80.00 66.67 65.56 rCBF + V e RF 76.67 ± 7.16 63.16 61.54 80.00 69.57 61.11 rCBV + rCBF + V e RF 76.67 ± 7.16 63.16 61.54 80.00 69.57 68.89 Abbreviations: CV, cross-validation; F1, F1-score; AUC, area under the curve; RF, random forest; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbor; rCBV, relative cerebral blood volume; rCBF, relative cerebral blood flow; V e , extravascular extracellular space volume fraction. Among single-parameter models, V e demonstrated the strongest cross-validation performance (RF: 74.44 ± 8.90% CV accuracy), while rCBV achieved the highest test AUC (RF: 74.44%). All three single-parameter models achieved identical test accuracy of 68.42% with consistent recall of 80.00%, indicating reliable sensitivity for DMG detection. Two-parameter combinations demonstrated improved test accuracy compared to single parameters. The rCBV and V e combination achieved the highest CV accuracy (RF: 76.67% ± 7.16%) among two-parameter models but moderate test accuracy (63.16%). Whereas the combination of rCBV and rCBF achieved the highest overall test accuracy of 73.68% using KNN, with precision of 72.73% and F1-score of 76.19%, though CV accuracy was 69.44% ± 10.54%. The three-parameter combination (rCBV + rCBF + V e ) using RF achieved the highest CV accuracy overall (76.67% ± 7.16%) with relatively low standard deviation, indicating stable CV performance with improved AUC (68.89%) compared to some two-parameter combinations. 4 Discussion This study comprehensively evaluated both individual DCE-MRI perfusion parameters and ML-based multi-parametric approaches for differentiating DMG from mGBM in children and young adults under 40 years of age. Our results demonstrate that individual parameters, particularly rCBF, rCBV, and V e , achieved robust differentiation (Table 3). ML classifiers, through systematic evaluation of all possible feature combinations from these three perfusion parameters, demonstrated that carefully selected combinations can provide clinically useful discrimination. The distinct perfusion profiles observed, with mGBM exhibiting the highest parametric values compared to DMG, reflect the underlying molecular and pathophysiological characteristics of these tumor entities [4,7]. The molecular basis for these perfusion differences is well-established in the literature. mGBM is characterized by EGFR amplification, TERTp mutations, and the upregulation of vascular endothelial growth factor (VEGF), promoting robust angiogenesis and vascular permeability, [4,7,29]. In contrast, H3K27-altered DMG exhibits distinct epigenetic dysregulation affecting cellular differentiation and vascular development, resulting in less aggressive angiogenic activity [4]. The significantly lower perfusion values in DMG compared to mGBM quantitatively validate this molecular distinction at the imaging level [17,18]. The clinical imperative for non-invasive differentiation is emphasized by the inherent challenges of tissue diagnosis in these tumors. While immunohistochemistry for H3 K27M mutation remains the gold standard for definitive DMG diagnosis [2,4], obtaining adequate tissue from eloquent midline structures presents formidable obstacles. Stereotactic biopsy of thalamic and brainstem lesions carries substantial risks including hemorrhage, neurological deficits, and procedure-related mortality, with complication rates reaching 10-15% in some series [12,13]. In patients with critical tumor locations or significant medical comorbidities, surgical sampling may be absolutely contraindicated. Thus, differentiation from other midline HGG tumors remains challenging yet clinically crucial (Figure 3) [12,14,17,18]. Beyond these molecular and clinical challenges, understanding the demographic context of these tumors is essential for appreciating the significance of our findings (Table 2). DMG with H3 K27-alteration has a bimodal age distribution, affecting primarily children/adolescents and young adults. Previous studies on DMGs have often focused on pediatric populations or included very wide age ranges [3,12,13]. However, our study extends previous work by specifically focusing on DMG differentiation in a carefully selected pediatric and young adult population (age < 40 years), addressing an important gap in the literature. By targeting this relatively homogeneous age group, we minimized confounding effects of age-related vascular changes while maintaining clinical relevance. Given the distinct characteristics of these tumors, accurate differentiation among midline HGG subtypes has important therapeutic implications for treatment planning, prognostication, and clinical trial eligibility [3,5,10]. Current treatment strategies differ substantially between these entities. mGBM typically receive maximal safe resection followed by chemoradiation [42,43], while DMGs may require alternative therapeutic approaches including targeted therapies, and clinical trial enrollment [8,11]. This therapeutic divergence underscores the importance of accurate preoperative characterization [4,5]. Conventional MRI remains the foundation of brain tumor imaging but has well-recognized limitations in distinguishing between DMG and other HGG subtypes as all typically appear as heterogeneously enhancing masses with surrounding edema and variable necrosis (Figure 3) [16,17,34]. While some features may suggest specific diagnoses, such as the strict midline location of DMGs, there is a substantial overlap in their imaging appearances, particularly between mGBMs and DMGs [17,18]. Our study indicates that advanced perfusion parameters can complement conventional imaging to improve diagnostic accuracy. Quantitative perfusion parameters effectively address this diagnostic challenge. Our findings show that both individual parameters and ML classifiers provide significant discriminatory power for this critical distinction. Specifically, rCBF, rCBV, and V e achieved AUC values of 75.31%, 71.25%, and 70.62%, respectively (Table 3), while ML classifiers achieved the highest CV accuracy of 76.67% (RF with rCBF+V e , and RF with rCBV+rCBF+V e ) and highest test accuracy of 73.68% (KNN with rCBV+rCBF) (Table 4). From a clinical implementation perspective, the findings have several important implications. First, quantitative DCE-MRI parameters can complement conventional MRI to improve diagnostic accuracy when molecular testing is unavailable or when biopsy is contraindicated due to eloquent tumor locations and associated procedural risks. Second, the automated analysis pipeline including LBV segmentation, feature extraction, and ML classification could potentially be integrated into clinical workflows to provide objective, reproducible tumor characterization. Third, the identification of key discriminatory features (rCBF, rCBV, V e ) suggests that simplified imaging protocols focusing on these parameters might be sufficient for clinical decision-making, potentially reducing computational burden. Finally, this approach ensures minimal risk of missing DMG cases that might benefit from targeted therapies or eligibility for clinical trial enrollment. 4.1 Limitations This study also has some limitations. First, the retrospective single-institution design may limit generalizability. Second, despite covering a six-year period (2018-2024), the sample size remains modest, particularly for the test set (19 cases: 9 mGBM, 10 DMG), which may affect the reliability of ML model performance estimates. Third, DMG patients are relatively rare and histopathological confirmation was challenging due to eloquent tumor locations. Fourth, our analysis incorporated data classified under both the 2016 and 2021 WHO classification systems, which might introduce classification inconsistencies [1,2]. Fifth, we focused exclusively on DCE-MRI perfusion parameters and did not integrate other advanced imaging modalities (diffusion-weighted imaging, susceptibility imaging, MR spectroscopy) or radiomics features, which might further improve classification accuracy. 4.2 Future Directions Future research should focus on validation of results of this study on large multicenter data. Integration of perfusion metrics with other advanced imaging features (diffusion parameters, spectroscopic metabolites, radiomics features) in comprehensive multimodal ML models may provide superior diagnostic performance. Deep learning approaches that can automatically extract relevant features from raw parametric maps without manual feature engineering represent a promising direction. 5 Conclusion This study evaluated the potential of quantitative DCE-MRI perfusion metrics and ML classifiers for the non-invasive differentiation of DMG from mGBM in children and young adults. DMG demonstrated significantly lower perfusion values compared to mGBM, consistent with their distinct biological profiles. Individual perfusion parameters, particularly rCBF, rCBV and V e achieved clinically useful discrimination and systematic evaluation of multi-parametric ML approaches, particularly RF provided incremental improvements. The interpretation of perfusion imaging in this study carries important implications for improving diagnostic accuracy in neuro-oncology, particularly when molecular testing through tissue acquisition is unavailable or contraindicated due to procedural risks. Abbreviations GBM, glioblastoma; mGBM, midline glioblastoma; DMG, diffuse midline glioma; HGG, high-grade glioma; MRI, magnetic resonance imaging; T1-W, T1-weighted; T2-W, T2-weighted; PD-W, proton density-weighted; FLAIR, fluid-attenuated inversion recovery; T1-Gd, postcontrast T1-W; DCE-MRI, dynamic contrast-enhanced magnetic resonance imaging; TR, repetition time; TE, echo time; FOV, field of view; PL, piecewise linear; ETM, extended Toft's model; rCBV, relative cerebral blood volume; rCBF, relative cerebral blood flow; Slope-2, washout slope; K trans , volume transfer constant; V e , extravascular extracellular space volume fraction; V p , blood plasma volume fraction; AI, artificial intelligence; LBV, large blood vessel; ROC, receiver operating characteristic; ML, machine learning; RF, random forest; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbors; CV, cross validation. Declarations Author Contribution Conceptualization: A.K., R.K.G., A.S.; Data curation: A.K.; Formal analysis: A.K.; Investigation: A.K.; Methodology: A.K., R.K.G., A.S.; Validation: A.K., R.K.G., A.S., S.A., R.P., S.V.; Visualization: A.K.; Writing – original draft: A.K.; Writing – review and editing: A.K., R.K.G., A.S., S.A., R.P., S.V.; Resources: R.K.G., S.A., R.P., S.V.; Supervision R.K.G., A.S.; Funding acquisition: A.S.; Project administration: A.S. Acknowledgement The authors thank Mr. Rakesh Kumar Singh from Fortis Memorial Research Institute, Gurugram, India for data acquisition and the MedImg Lab members for their constant guidance and support. Data Availability The code and a sample of anonymized test data can be accessed after sending an appropriate email request to the corresponding author. Ethics and Consent to Participate Declarations This study was approved by the institutional ethics committee of XYZ Hospital (name anonymized per journal requirements), and the requirement for informed consent was waived by the committee. Funding Declaration This work was supported by the Science and Engineering Research Board, Department of Science & Technology (project number: CRG/2019/005032) and Indian Council of Medical Research, Government of India (Project number: CAR-2024-01-000187). References Louis DN, Perry A, Reifenberger G, von Deimling A, Figarella-Branger D, Cavenee WK, Ohgaki H, Wiestler OD, Kleihues P, Ellison DW (2016) The 2016 World Health Organization Classification of tumors of the central nervous system: a summary. Acta Neuropathol 131:803–820. https://doi.org/10.1007/s00401-016-1545-1 Louis DN, Perry A, Wesseling P, Brat DJ, Cree IA, Figarella-Branger D, Hawkins C, Ng HK, Pfister SM, Reifenberger G, Soffietti R, von Deimling A, Ellison DW (2021) The 2021 WHO classification of tumors of the central nervous system: a summary. Neuro Oncol 23:1231–1251. https://doi.org/10.1093/neuonc/noab106 Zheng L, Gong J, Yu T, Zou Y, Zhang M, Nie L, Chen X, Yue Q, Liu Y, Mao Q, Zhou Q, Chen N (2022) Diffuse midline gliomas with histone H3 K27M mutation in adults and children: a retrospective series of 164 cases. Am J Surg Pathol 46:863–871. https://doi.org/10.1097/PAS.0000000000001897 Park YW, Vollmuth P, Foltyn-Dumitru M, Sahm F, Ahn SS, Chang JH, Kim SH (2023) The 2021 WHO classification for gliomas and implications on imaging diagnosis: Part 2—summary of imaging findings on pediatric-type diffuse high-grade gliomas, pediatric-type diffuse low-grade gliomas, and circumscribed astrocytic gliomas. J Magn Reson Imaging 58:690–708. https://doi.org/10.1002/jmri.28543 Al Sharie S, Abu Laban D, Al-Hussaini M (2023) Decoding diffuse midline gliomas: a comprehensive review of pathogenesis, diagnosis and treatment. Cancers (Basel) 15:4869. https://doi.org/10.3390/cancers15194869 Lovibond S, Gewirtz AN, Pasquini L, Krebs S, Graham MS (2023) The promise of metabolic imaging in diffuse midline glioma. Neoplasia 39:100896. https://doi.org/10.1016/j.neo.2023.100896 Park YW, Vollmuth P, Foltyn-Dumitru M, Sahm F, Ahn SS, Chang JH, Kim SH (2023) The 2021 WHO classification for gliomas and implications on imaging diagnosis: Part 1—key points of the fifth edition and summary of imaging findings on adult-type diffuse gliomas. J Magn Reson Imaging 58:677–689. https://doi.org/10.1002/jmri.28743 Akdemir EY, Odia Y, Hall MD, Mehta MP, Kotecha R (2024) An update on H3K27M-altered diffuse midline glioma: diagnostic and therapeutic challenges in clinical practice. Pract Radiat Oncol 14:443–451. https://doi.org/10.1016/j.prro.2024.04.013 Joshi S, Wu YK, Xu YK, Liu XM, Zhang H (2022) Comparative study of diffuse midline glioma and glioblastoma: Magnetic Resonance Imaging (MRI) in the characteristics and demography. Nepal J Neurosci 19:32–36. https://doi.org/10.3126/njn.v19i2.43724 Pachocki CJ, Hol EM (2022) Current perspectives on diffuse midline glioma and a different role for the immune microenvironment compared to glioblastoma. J Neuroinflammation 19:212. https://doi.org/10.1186/s12974-022-02630-8 Chi AS, Tarapore RS, Hall MD, et al (2019) Pediatric and adult H3 K27M-mutant diffuse midline glioma treated with the selective DRD2 antagonist ONC201. J Neurooncol 145:97–105. https://doi.org/10.1007/s11060-019-03271-3 Sim Y, Choi SH, Lee N, Park YW, Ahn SS, Chang JH, Kim SH, Lee SK (2024Clinical, qualitative imaging biomarkers, and tumor oxygenation imaging biomarkers for differentiation of midline-located IDH wild-type glioblastomas and H3 K27-altered diffuse midline gliomas in adults. Eur J Radiol 173:111384. https://doi.org/10.1016/j.ejrad.2024.111384 Wiśniewski K, Ghaly A, Drummond K, Fahlström A (2023) H3 K27M-altered diffuse midline gliomas: a review. Indian J Neurosurg 12:104–115. https://doi.org/10.1055/s-0043-1771192 Lv K, Chen H, Cao X, Du P, Chen J, Liu X, Zhu L, Geng D, Zhang J (2023) Development and validation of a machine learning algorithm for predicting diffuse midline glioma, H3 K27–altered, H3 K27 wild-type high-grade glioma, and primary CNS lymphoma of the brain midline in adults. J Neurosurg 39:393–401. https://doi.org/10.3171/2022.11.JNS221544 Chauhan RS, Kulanthaivelu K, Kathrani N, Kotwal A, Bhat MD, Saini J, Srinivas D (2021) Prediction of H3K27M mutation status of diffuse midline gliomas using MRI features. J Neuroimaging 31:1201–1210. https://doi.org/10.1111/jon.12905 Chauhan RS, Kulanthaivelu K, Kathrani N, et al (2021) Can conventional MRI features predict H3K27M mutation status of diffuse midline gliomas? Res Square Preprint. https://doi.org/10.21203/rs.3.rs-360529/v1 Banan R, Akbarian A, Samii M, Samii A, Bertalanffy H, Lehmann U, Hartmann C, Brüning R (2021) Diffuse midline gliomas, H3 K27M-mutant are associated with less peritumoral edema and contrast enhancement in comparison to glioblastomas, H3 K27M-wildtype of midline structures. PLoS One 16:e0249647. https://doi.org/10.1371/journal.pone.0249647 Raab P, Banan R, Akbarian A, et al (2022) Differences in the MRI Signature and ADC Values of Diffuse Midline Gliomas with H3 K27M Mutation Compared to Midline Glioblastomas. Cancers (Basel) 14:1397. https://doi.org/10.3390/cancers14061397 Kim M, Jung SY, Park JE, Jo Y, Park SY, Nam SJ, Kim JH, Kim HS (2020) Diffusion- and perfusion-weighted MRI radiomics model may predict isocitrate dehydrogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade glioma. Eur Radiol 30:2142–2151. https://doi.org/10.1007/s00330-019-06548-3 Guo W, She D, Xing Z, Lin X, Wang F, Song Y, Cao D (2022) Multiparametric MRI-Based Radiomics Model for Predicting H3 K27M Mutant Status in Diffuse Midline Glioma: A Comparative Study Across Different Sequences and Machine Learning Techniques, Front Oncol 12:796583. https://doi.org/10.3389/FONC.2022.796583 Kurokawa R, Baba A, Kurokawa M, et al (2022) Perfusion and diffusion-weighted imaging parameters: Comparison between pre-and postbiopsy MRI for high-grade glioma. Medicine 101:e30183. https://doi.org/10.1097/MD.0000000000030183 Rameh V, Vajapeyam S, Ziaei A, et al (2023) Correlation between Multiparametric MR Imaging and Molecular Genetics in Pontine Pediatric High-Grade Glioma. AJNR Am J Neuroradiol 44:833–840. https://doi.org/10.3174/ajnr.A7910 Szychot E, Youssef A, Ganeshan B, et al (2021) Predicting outcome in childhood diffuse midline gliomas using magnetic resonance imaging based texture analysis. J Neuroradiol 48:243–247. https://doi.org/10.1016/j.neurad.2020.02.005 Kathrani N, Chauhan RS, Kotwal A, et al (2022) Diffusion and perfusion imaging biomarkers of H3 K27M mutation status in diffuse midline gliomas. Neuroradiology 64:1519–1528. https://doi.org/10.1007/s00234-021-02857-x Wu C, Zheng H, Li J, et al (2022) MRI-based radiomics signature and clinical factor for predicting H3K27M mutation in pediatric high-grade gliomas located in the midline of the brain. Eur Radiol 32:1813–1822. https://doi.org/10.1007/s00330-021-08234-9 Kurokawa R, Kurokawa M, Baba A, et al (2023) Dynamic susceptibility contrast perfusion-weighted and diffusion-weighted magnetic resonance imaging findings in pilocytic astrocytoma and H3.3 and H3.1 variant diffuse midline glioma, H3K27-altered. PLoS One 18:e0288412. https://doi.org/10.1371/journal.pone.0288412 Cramer S, Popa L, Haley S, et al (2020) Path-04. the blood-brain barrier in diffuse midline glioma and its implications for drug delivery. Neuro Oncol 22:ii164. https://doi.org/10.1093/neuonc/noaa215.686 Tofts PS, Kermode AG (1991) measurement of the blood-brain barrier permeability and leakage space using dynamic MR imaging. 1. fundamental concepts. Magn Reson Med 17:357–367. https://doi.org/10.1002/mrm.1910170223 Cuenod CA, Balvay D (2013) Perfusion and vascular permeability: basic concepts and measurement in DCE-CT and DCE-MRI. Diagn Interv Imaging 94:1187–1204. https://doi.org/10.1016/j.diii.2013.10.010 Jain R, Griffith B, Alotaibi F, et al (2015) Glioma angiogenesis and perfusion imaging: Understanding the relationship between tumor blood volume and leakiness with increasing glioma grade. AJNR Am J Neuroradiol 36:2030–2035. https://doi.org/10.3174/ajnr.A4405 Chen YJ, Chu WC, Pu YS, et al (2012) Washout gradient in dynamic contrast-enhanced MRI is associated with tumor aggressiveness of prostate cancer. J Magn Reson Imaging 36:912–919. https://doi.org/10.1002/jmri.23723 Gupta PK, Saini J, Sahoo P, et al (2017) Role of Dynamic Contrast-Enhanced Perfusion Magnetic Resonance Imaging in Grading of Pediatric Brain Tumors on 3T. Pediatr Neurosurg 52:298–305. https://doi.org/10.1159/000479283 Law M, Yang S, Babb JS, et al (2004) Comparison of Cerebral Blood Volume and Vascular Permeability from Dynamic Susceptibility Contrast-Enhanced Perfusion MR Imaging with Glioma Grade. AJNR Am J Neuroradiol 25:746–755 Saini J, Gupta PK, Sahoo P, et al (2018) Differentiation of grade II/III and grade IV glioma by combining “T1 contrast-enhanced brain perfusion imaging” and susceptibility-weighted quantitative imaging. Neuroradiology 60:43–50. https://doi.org/10.1007/s00234-017-1942-8 Sengupta A, Agarwal S, Gupta PK, et al (2018) On differentiation between vasogenic edema and non-enhancing tumor in high-grade glioma patients using a support vector machine classifier based upon pre and post-surgery MRI images. Eur J Radiol 106:199–208. https://doi.org/10.1016/j.ejrad.2018.07.018 Singh A, Rathore RKS, Haris M, et al (2009) Improved bolus arrival time and arterial input function estimation for tracer kinetic analysis in DCE-MRI. J Magn Reson Imaging 29:166–176. https://doi.org/10.1002/jmri.21624 Singh A, Haris M, Rathore D, et al (2007) Quantification of physiological and hemodynamic indices using T1 dynamic contrast- enhanced MRI in intracranial mass lesions. J Magn Reson Imaging 26:871–880. https://doi.org/10.1002/jmri.21080 Ashburner J, Barnes G, Chen CC, et al (2021) SPM12 Manual The FIL Methods Group (and honorary members), London. https://www.fil.ion.ucl.ac.uk/spm/ Sahoo P, Gupta RK, Gupta PK, et al (2017 Diagnostic accuracy of automatic normalization of CBV in glioma grading using T1- weighted DCE-MRI. Magn Reson Imaging 44:32–37. https://doi.org/10.1016/j.mri.2017.08.003 Maurya S, Yadav VK, Agarwal S, Singh A (2021) Brain Tumor Segmentation in mpMRI scans (BraTS-2021) using models based on U-Net Architecture, In: International MICCAI Brainlesion Workshop. LNCS 12963:312–323. https://doi.org/10.1007/978-3-031-09002-8_28 Kesari A, Maurya S, Sheikh MT, Gupta RK, Singh A (2025) Large blood vessel segmentation in quantitative DCE-MRI of brain tumors: A Swin UNETR approach. Magn Reson Imaging 118:110342. https://doi.org/10.1016/j.mri.2025.110342 Majewska P, Ioannidis S, Raza MH, et al (2017) Postprogression survival in patients with glioblastoma treated with concurrent chemoradiotherapy: a routine care cohort study. CNS Oncol 6:307. https://doi.org/10.2217/cns-2017-0001 Stupp R, Mason WP, van den Bent MJ, et al (2005) Radiotherapy plus Concomitant and Adjuvant Temozolomide for Glioblastoma. N Engl J Med 352:987–996. https://doi.org/10.1056/NEJMoa043330 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8481592","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":573435675,"identity":"8844d1d5-46b4-4708-9134-105656579a28","order_by":0,"name":"Anshika Kesari","email":"","orcid":"","institution":"Indian Institute of Technology Delhi","correspondingAuthor":false,"prefix":"","firstName":"Anshika","middleName":"","lastName":"Kesari","suffix":""},{"id":573435676,"identity":"6f512203-1918-4b75-a5f6-1675f19831ee","order_by":1,"name":"Rakesh Kumar Gupta","email":"","orcid":"","institution":"Fortis Memorial Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Rakesh","middleName":"Kumar","lastName":"Gupta","suffix":""},{"id":573435677,"identity":"597a5e24-b706-41e8-8ce8-2262863775fe","order_by":2,"name":"Sunita Ahlawat","email":"","orcid":"","institution":"Fortis Memorial Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Sunita","middleName":"","lastName":"Ahlawat","suffix":""},{"id":573435679,"identity":"d5dce15c-f8db-464b-8399-5f9623d742c6","order_by":3,"name":"Rana Patir","email":"","orcid":"","institution":"Fortis Memorial Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Rana","middleName":"","lastName":"Patir","suffix":""},{"id":573435680,"identity":"9c90a253-2826-4f38-be31-4813691f0276","order_by":4,"name":"Sandeep Vaishya","email":"","orcid":"","institution":"Fortis Memorial Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Sandeep","middleName":"","lastName":"Vaishya","suffix":""},{"id":573435682,"identity":"318d3875-2c27-4067-a040-2ff67ab6e4c8","order_by":5,"name":"Anup Singh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvklEQVRIiWNgGAWjYPACGx42BgYDEIuHWC1pPGxsJGo5zMAA1UIY8M/uMXv4M+e8DJ988waGHzUMMuaEtEjcOWNuzLvtNtBhbAWMPccYeCwbCOm5kWMmzQjWwmPAwNvAwGNwgIAOeaAWyZ/bzoG1MP4lRosBUIsE77YDYC3MRNlieCOtHOiXZKCWtILDMsckCGuRu5G87eHPbXb28s2HNz58U2NjT1ALAyhCYACoWIKwehQto2AUjIJRMAqwAgAPdjRGiHmZngAAAABJRU5ErkJggg==","orcid":"","institution":"Indian Institute of Technology Delhi","correspondingAuthor":true,"prefix":"","firstName":"Anup","middleName":"","lastName":"Singh","suffix":""}],"badges":[],"createdAt":"2025-12-30 13:08:26","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8481592/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8481592/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100399844,"identity":"05e24a4b-f26d-4911-b7f0-49b8f5b70e69","added_by":"auto","created_at":"2026-01-16 11:57:39","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1577906,"visible":true,"origin":"","legend":"","description":"","filename":"FinalManuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/aa3fe70aecd81b61202bf093.docx"},{"id":100400440,"identity":"35661c45-8a8e-45c4-9e26-c4203f4493c8","added_by":"auto","created_at":"2026-01-16 11:58:09","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8257,"visible":true,"origin":"","legend":"","description":"","filename":"bbbf12c8c5eb45ed925ed18cccc2230f.json","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/c8037268e55724b9d9dc8596.json"},{"id":100401045,"identity":"085874bb-e416-41ac-96a2-73367bab9b1e","added_by":"auto","created_at":"2026-01-16 11:58:40","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":138818,"visible":true,"origin":"","legend":"","description":"","filename":"bbbf12c8c5eb45ed925ed18cccc2230f1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/11d811db500f627357baebc6.xml"},{"id":100401254,"identity":"9f8ff260-d29d-4dee-bf5a-b3931983041f","added_by":"auto","created_at":"2026-01-16 11:58:49","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":27268,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/c8b57647851e8b05de1e1970.png"},{"id":100400445,"identity":"36aab94c-4be9-44a8-929d-3f75b2192c6c","added_by":"auto","created_at":"2026-01-16 11:58:10","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":73676,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/e391268d637e585ef00acea4.png"},{"id":100401433,"identity":"47429fcc-816a-4d14-83b5-f25ad5df9633","added_by":"auto","created_at":"2026-01-16 11:58:54","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":766212,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/09776a9ca4fb9a850bdee6ea.png"},{"id":100400962,"identity":"be26c32c-7cf7-454b-b9a7-dbc0895f5def","added_by":"auto","created_at":"2026-01-16 11:58:36","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":131453,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/5485372aac8691bdb246000f.png"},{"id":100401008,"identity":"9e011338-01f7-406c-8cd4-227e187c80a7","added_by":"auto","created_at":"2026-01-16 11:58:39","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128390,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/0097c723d1622fce0353f928.png"},{"id":100400443,"identity":"1ea29e58-62ae-42d5-9fd0-010961811b6e","added_by":"auto","created_at":"2026-01-16 11:58:09","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15202,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/f5e896ea6a487c39e55bf866.png"},{"id":100401105,"identity":"56ef74b4-e6c9-43c3-8482-ea5c61916502","added_by":"auto","created_at":"2026-01-16 11:58:42","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70217,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/5669908c09ec76fba4444ec8.png"},{"id":100401430,"identity":"b1339002-0adf-4790-bf9e-f26c1177bfaf","added_by":"auto","created_at":"2026-01-16 11:58:54","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":132298,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/40aea8ac50958c50279de671.png"},{"id":100401286,"identity":"dda7fc36-8739-405a-99e4-39c709f93af4","added_by":"auto","created_at":"2026-01-16 11:58:50","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28089,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/ecc402959ede231a59542bae.png"},{"id":100399891,"identity":"a2a929bb-3220-4d2c-9233-f3851f42ac08","added_by":"auto","created_at":"2026-01-16 11:57:45","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":27746,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/0ddfa120c04722515a253d7b.png"},{"id":100399854,"identity":"a9c811bf-a80b-41b0-9ed9-dbf0a66f477b","added_by":"auto","created_at":"2026-01-16 11:57:39","extension":"xml","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137789,"visible":true,"origin":"","legend":"","description":"","filename":"bbbf12c8c5eb45ed925ed18cccc2230f1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/3bb732f9cf070812ac3c9910.xml"},{"id":100400437,"identity":"f3fe9c12-71ae-4a58-beed-39b379e42f0a","added_by":"auto","created_at":"2026-01-16 11:58:09","extension":"html","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":149046,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/2cadbcd18f1392677993c150.html"},{"id":100400950,"identity":"cf18efca-f4c1-4c82-a846-28e9de16c07e","added_by":"auto","created_at":"2026-01-16 11:58:36","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":58745,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram illustrating patient selection for the study.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/cee0288f06bafd7549addaa2.png"},{"id":100401299,"identity":"f44f6460-f98d-4969-9d60-705beef321dc","added_by":"auto","created_at":"2026-01-16 11:58:51","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":174425,"visible":true,"origin":"","legend":"\u003cp\u003eMethodology flow diagram depicting the complete processing pipeline: perfusion processing, tumor and large blood vessel (LBV) segmentation for generating the non-vascular tumor region, applying the statistics on the generated quantitative parameters of non-vascular tumor region, and applying machine learning for differentiating among the two midline glioma subtypes. Abbreviations: mGBM, midline glioblastoma; DMG, diffuse midline glioma; T1-W, T1-weighted; T2-W, T2-weighted; PD-W, proton density-weighted; FLAIR, fluid-attenuated inversion recovery; T1-Gd, postcontrast T1-W; DCE-MRI, dynamic contrast-enhanced magnetic resonance imaging; PL, piecewise linear; ETM, extended Toft's model; rCBV, relative cerebral blood volume; rCBF, relative cerebral blood flow; Slope-2, washout slope; K\u003csup\u003etrans\u003c/sup\u003e, volume transfer constant; V\u003csub\u003ee\u003c/sub\u003e, extravascular extracellular space volume fraction; V\u003csub\u003ep\u003c/sub\u003e, blood plasma volume fraction; AI, artificial intelligence; LBV, large blood vessel; ROC, receiver operating characteristic; ML, machine learning; RF, random forest; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbors; CV, cross validation.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/ff8b0251b4f72d5d4477688e.png"},{"id":100400292,"identity":"11c608c0-6054-434b-9bcd-aa94a8151b1f","added_by":"auto","created_at":"2026-01-16 11:58:03","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":718453,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative MRI images of midline gliomas demonstrating the diagnostic challenge of conventional imaging. Case 1 shows contrast-enhancing tumors, including mGBM (1\u003csup\u003est\u003c/sup\u003e row), and DMG (2\u003csup\u003end\u003c/sup\u003e row), with axial FLAIR (1\u003csup\u003est\u003c/sup\u003e and 5\u003csup\u003eth\u003c/sup\u003e column) and post-contrast T1-weighted (T1-Gd) (2\u003csup\u003end\u003c/sup\u003e and 6\u003csup\u003eth\u003c/sup\u003e column) images showing typical imaging characteristics and their corresponding rCBV maps (3\u003csup\u003erd\u003c/sup\u003e and 7\u003csup\u003eth\u003c/sup\u003e column) and rCBV maps without LBVs (4\u003csup\u003eth\u003c/sup\u003e and 8\u003csup\u003eth\u003c/sup\u003e column). Case 2 shows non-enhancing variants with the same sequence display.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/99e8c6e701bbeef7be62902f.png"},{"id":100399614,"identity":"209e94be-b565-4d00-9436-3a9ad8e43fbe","added_by":"auto","created_at":"2026-01-16 11:57:19","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":153680,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of perfusion parameters between midline gliomas. (a) Box and whisker plots showing minimum, first quartile (Q1), mean (red cross), median, third quartile (Q3), maximum, and outliers of 95\u003csup\u003eth\u003c/sup\u003e percentile values for rCBV, rCBF, and V\u003csub\u003ee\u003c/sub\u003e across the DMG and mGBM. (b) Violin plots illustrating the distribution and density of these parameters, with significant differences observed between tumor types. Abbreviations: mGBM, midline glioblastoma; DMG, diffuse midline glioma; rCBV, relative cerebral blood volume; rCBF, relative cerebral blood flow; V\u003csub\u003ee\u003c/sub\u003e, extravascular extracellular space volume fraction.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/5abf1116313115acab4ab70e.png"},{"id":100401117,"identity":"a43e4f53-5e7f-4df3-9899-a02fa2783198","added_by":"auto","created_at":"2026-01-16 11:58:42","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":162009,"visible":true,"origin":"","legend":"\u003cp\u003eROC curve analysis and confusion matrices for differentiation between midline gliomas using multiple perfusion parameters. (a) ROC curves of all six perfusion parameters, and (b) confusion matrices of top three perfusion parameters (rCBV, rCBF, V\u003csub\u003ee\u003c/sub\u003e) comparing the discriminatory performance of 95\u003csup\u003eth\u003c/sup\u003e percentile values for distinguishing between mGBM and DMG. Abbreviations: ROC, receiver operating characteristic; mGBM, midline glioblastoma; DMG, diffuse midline glioma; rCBV, relative cerebral blood volume; rCBF, relative cerebral blood flow; Slope2, washout slope; K\u003csup\u003etrans\u003c/sup\u003e, volume transfer constant; V\u003csub\u003ee\u003c/sub\u003e, extravascular extracellular space volume fraction; V\u003csub\u003ep\u003c/sub\u003e, blood plasma volume fraction.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/098ec8ea409b10008f5abe18.png"},{"id":100412350,"identity":"ad3fb1ed-6df8-4e91-acf6-93c037b27864","added_by":"auto","created_at":"2026-01-16 13:14:17","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1922454,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8481592/v1/f94a084f-df15-45cd-b407-0fa090f9011f.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Non-Invasive Differentiation of Diffuse Midline Glioma and Midline Glioblastoma Using DCE-MRI Perfusion Parameters and Machine Learning Classification in Pediatric and Young Adult Patients","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eH3 K27M-mutant diffuse midline glioma (DMG) was recognized as a distinct brain tumor type in the 2016 World Health Organization (WHO) Classification of Tumors of the Central Nervous System. This classification was updated in 2021, renaming the entity as H3 K27-altered DMG defined by mutations affecting the histone H3 protein resulting in K27M amino acid substitution [1,2]. This molecular definition supersedes histological grading, with all H3 K27-altered DMGs considered WHO grade 4, irrespective of histological features or location. Although DMGs have been classified as pediatric-type diffuse high-grade gliomas (HGGs), they are also frequently diagnosed in young adults and occasionally in older patients [3]. These tumors primarily arise in central midline structures including the thalamus, brainstem, and spinal cord [3] and are associated with poor prognosis [4\u0026ndash;6]. However, accurate diagnosis of DMG can be challenging due to imaging similarities with other aggressive midline tumors, particularly midline glioblastoma (mGBM), despite their distinct molecular profiles and treatment requirements.\u003c/p\u003e\n\u003cp\u003eAccurate differentiation between DMG and mGBM has critical therapeutic and prognostic implications despite their similarly aggressive nature. Both entities are WHO grade 4 tumors with comparable median survival of approximately 12-18 months [2,4,7]. However, they require fundamentally different treatment strategies with significant implications for patient management and clinical outcomes. Standard treatment for mGBM consists of maximal safe resection followed by concurrent temozolomide chemotherapy and radiotherapy (Stupp protocol), with adjuvant temozolomide maintenance [8\u0026ndash;10].\u0026nbsp;This approach demonstrates survival benefit in appropriately selected patients. In contrast, DMG management emphasizes radiation therapy as the primary treatment modality, as extensive surgical resection is often precluded by eloquent tumor locations [5,8,9]. Importantly, emerging targeted therapies including histone deacetylase (HDAC) inhibitors and ONC201 show specific promise for H3 K27-altered DMG, making accurate preoperative identification essential for clinical trial eligibility and personalized treatment planning [5,10,11]. Given these therapeutic implications, reliable preoperative distinction between these entities is essential.\u003c/p\u003e\n\u003cp\u003eThe gold standard for definitive differentiation relies on molecular testing for H3 K27-alteration through immunohistochemistry or genetic sequencing of tumor tissue [2,4]. However, obtaining adequate tissue for analysis presents substantial clinical challenges due to the eloquent midline locations of these tumors [9,12\u0026ndash;14]. Stereotactic biopsy in these regions carries considerable risks of neurological complications, hemorrhage, and procedure-related morbidity, and in many cases, surgical intervention is contraindicated due to critical anatomical location or patient frailty [9,12,13]. Furthermore, conventional magnetic resonance imaging (MRI) sequences, including T1-weighted (T1-W), T2-weighted (T2-W), and fluid-attenuated recovery (FLAIR) images, demonstrate significant limitations in distinguishing these entities. Both tumor types characteristically present as heterogeneously enhancing masses with surrounding edema and variable necrosis on conventional imaging [9,15\u0026ndash;17]. While strict midline location may suggest DMG, substantial overlap in imaging appearances between mGBM and DMG complicates reliable non-invasive differentiation [12\u0026ndash;14,17,18], underscoring the need for advanced imaging biomarkers. To address this clinical need, advanced MRI techniques beyond conventional sequences have emerged as promising non-invasive tools.\u003c/p\u003e\n\u003cp\u003eMR diffusion-weighted imaging (DWI) provides cellularity information through apparent diffusion coefficient (ADC) mapping [18\u0026ndash;25]. MR perfusion-weighted imaging (PWI), including both dynamic susceptibility contrast (DSC) [12,19,24,26] and dynamic contrast-enhanced (DCE) MRI [27\u0026ndash;29], offers complementary information about tumor vascularity. Among these, DCE-MRI enables quantitative assessment of tissue vascularity and permeability by tracking the kinetics of contrast agent distribution. Quantitative perfusion parameters derived from DCE-MRI, including relative cerebral blood volume (rCBV), and relative cerebral blood flow (rCBF), K\u003csup\u003etrans\u003c/sup\u003e (volume transfer constant), K\u003csub\u003eep\u003c/sub\u003e (rate constant), V\u003csub\u003ee\u003c/sub\u003e (extravascular extracellular space volume fraction), and V\u003csub\u003ep\u003c/sub\u003e (plasma volume fraction),\u0026nbsp;provide insights into tumor angiogenesis and blood-brain barrier integrity [18,24,30\u0026ndash;33].\u003c/p\u003e\n\u003cp\u003ePrevious studies have examined various imaging approaches for DMG characterization, ranging from conventional MRI features [16,17,34] to advanced techniques including diffusion [18,20,22,24] and perfusion imaging [12,24,26]. Recently, classical machine learning (ML) methods utilizing perfusion features have been applied for tumor classification, showing that combining multiple DCE-MRI parameters can improve classification accuracy compared to single parameter approaches [14,20,35]. However, comprehensive studies directly comparing individual DCE-MRI quantitative perfusion parameters and systematic evaluation of multi-parametric ML classifiers specifically for distinguishing DMG from mGBM in children and young adults remain limited.\u003c/p\u003e\n\u003cp\u003eAddressing these gaps, the present study aims to evaluate the utility of quantitative DCE-MRI perfusion parameters using both individual parameters and multi-parametric MRI-based machine learning classifiers for distinguishing DMG from mGBM in children and young adults. These entities represent a particular diagnostic challenge due to their overlapping anatomical distribution and similar conventional imaging features, including contrast enhancement patterns and the presence of necrosis. We systematically assess the discriminatory performance of individual perfusion metrics and optimized ML combinations, and their relationship to molecular features, particularly H3 K27-alteration status.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cp\u003e2.1 Patient Population\u003c/p\u003e\n\u003cp\u003eThe dataset was acquired from the \u003cstrong\u003eXYZ\u003c/strong\u003e hospital. The retrospective study protocol was approved by the institutional ethics committee, and the requirement for informed consent was waived by the committee. Patients with HGG and MRI data between 2018-2024 were identified. The study applied strict inclusion criteria: (1) histo-pathologically confirmed diagnosis of DMG, or mGBM; (2) availability of preoperative DCE-MRI data along with conventional MRI; and (3) molecular testing results for IDH mutation or H3 K27-alteration (Figure 1). Patients were excluded beforehand if they had prior treatment, significant motion artefacts on MRI, incomplete clinical data or any other type of tumor.\u003c/p\u003e\n\u003cp\u003eIn the current study, 62 patients satisfied the patient selection criteria including age (\u0026lt; 40 years) criteria. Figure 1 shows the final cohort that met the inclusion and exclusion criteria (30 mGBMs, and 32 DMGs). Patient demographics, tumor location, and molecular characteristics were extracted from their records. Tumors were classified according to WHO Classification of Central Nervous System Tumors [1,2].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.2 MRI Acquisition Protocol\u003c/p\u003e\n\u003cp\u003eAll imaging studies were performed on a 3T whole-body MRI system (Ingenia, Philips Healthcare, Netherlands) using a 15-channel head coil. The MRI protocol included axial\u0026nbsp;T1-W turbo-spin echo (TSE), axial dual PD-T2-W (TSE), axial FLAIR followed by 3D DCE-MRI (T1-W perfusion MRI) using fast-field echo (FFE) and axial post contrast T1-W (T1-Gd).The contrast injection was administered after the acquisition of four baseline scans to establish pre-contrast signal intensity.\u003c/p\u003e\n\u003cp\u003eDCE-MRI acquisition included 32 dynamic acquisitions with a temporal resolution of 3.8 seconds. A gadolinium-based contrast agent Dotarem (Gadoterate Meglumine, Guerbet, France) was administered intravenously at a dose of 0.1 mmol/kg body weight at a rate of 2.5 ml/sec using a power injector, followed by a 20 ml saline flush. Acquisition parameters for each sequence are listed in Table 1.\u003c/p\u003e\n\u003cp\u003e2.3 Histopathological and Conventional MRI Analysis\u003c/p\u003e\n\u003cp\u003eThe conventional MRI features were assessed by an experienced neuroradiologist\u0026nbsp;with 38 years of experience in neuroimaging, along with\u0026nbsp;histopathological and molecular results. The following features were assessed: (1) tumor location; (2) patient age; (3)\u0026nbsp;H3K27-alteration status;\u0026nbsp;(4) presence of contrast enhancement; and (5) necrosis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1.\u0026nbsp;\u003c/strong\u003eDetailed protocol for MRI sequences used in the current study.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMRI Sequences\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTR/TE \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;(ms)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFlip Angle\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e(ϴ)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNo. of slices\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSlice Thickness (mm)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFOV \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;(mm\u003csup\u003e2\u003c/sup\u003e)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAcquisition Matrix\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eT1-W\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e360/10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e230 \u0026times; 230\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e244 x 237\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eT2-W\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e3500/90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e230 \u0026times; 230\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e244 x 237\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003ePD-W\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e3500/23.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e230 \u0026times; 230\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e244 x 237\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eFLAIR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e4800/340\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e250 \u0026times; 250\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e224 x 223\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eT1-Gd\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e700/35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e250 \u0026times; 250\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e280 x 278\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eDCE-MRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e6.38/3.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 84px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e230 \u0026times; 230\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e192 x 177\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAbbreviations: MRI, magnetic resonance imaging; TR, repetition time; TE, echo time; FOV, field of view; T1-W, T1-weighted; T2-W, T2-weighted; PD-W, proton density-weighted; FLAIR, fluid-attenuated inversion recovery; T1-Gd, postcontrast T1-W; DCE-MRI, dynamic contrast-enhanced magnetic resonance imaging.\u003c/p\u003e\n\u003cp\u003e2.4 DCE-MRI Data Processing\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe comprehensive pipeline of the whole process is shown in Figure 2. DCE-MRI data processing was performed for quantitative analysis using an in-house developed MATLAB-based software solution (R2021a, MathWorks Inc., Natick, MA, USA) [36,37]. The processing pipeline included the following steps:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(1) 3D rigid body registration of all images with respect to T1-W images using SPM12 (Statistical Parametric Mapping software) [38];\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(2) De-sculping, and segmentation of gray matter, white matter, and cerebrospinal fluid using SPM12 followed by spatial smoothing;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(3) Intensity scaling of structural MRI images (T1-W, T2-W and PD-W) to obtain true pixel signal intensity.\u003c/p\u003e\n\u003cp\u003e(4) Computing pre-contrast T1 maps using\u0026nbsp;T1-W and dual PD-T2-W.\u003c/p\u003e\n\u003cp\u003e(5) Conversion of signal intensity time curves to concentration time curves.\u003c/p\u003e\n\u003cp\u003e(6) Local arterial input function was estimated using the previously reported algorithm for reducing the partial volume effect [36].\u003c/p\u003e\n\u003cp\u003e(7) Concentration time curves were analyzed voxel-wise to compute various quantitative perfusion parameters using different mathematical models.\u003c/p\u003e\n\u003cp\u003eFor the quantification of perfusion parameters, first-pass analysis was performed using the first pass of the contrast bolus through the brain vasculature to compute hemodynamic parameters, including CBV and CBF [37]. \u0026nbsp;Piecewise linear model fitting of the complete concentration time curve was used for the estimation of bolus arrival time (BAT), time to peak during first pass of bolus (BETA), slopes of second (wash-in slope; Slope1) and third line segments (wash-out slope; Slope-2) [36]. The extended Tofts model was applied to estimate the pharmacokinetic parameters, including K\u003csup\u003etrans\u003c/sup\u003e, V\u003csub\u003ee\u003c/sub\u003e, and V\u003csub\u003ep\u003c/sub\u003e using the bidirectional transport of contrast agent between the intravascular and extravascular extracellular spaces [28,36].\u003c/p\u003e\n\u003cp\u003eHemodynamic parameters were normalized with respect to the average values of these parameters in contralateral normal-appearing white matter region to obtain their relative values (rCBV, rCBF), which compensates for potential variations in contrast administration and physiological factors across patients [39].\u003c/p\u003e\n\u003cp\u003e2.5 Tumor Segmentation\u003c/p\u003e\n\u003cp\u003eTumor segmentation was performed on FLAIR images using an artificial intelligence (AI) algorithm based on a U-Net architecture [40]. The initial automated segmentation was subsequently reviewed and, when necessary, manually refined by an experienced neuroradiologist. The resulting tumor masks encompassed the entire abnormal signal intensity on FLAIR images, including necrosis, enhancing and non-enhancing components, as well as peritumoral edema.\u003c/p\u003e\n\u003cp\u003e2.6 Large Blood Vessel (LBV) Segmentation\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor the segmentation of non-tumoral LBVs, we employed an automated approach based on the combined analysis of rCBV and Slope-2 maps derived from DCE-MRI [41]. The resulting LBV mask was then subtracted from the tumor mask to generate a refined tumor mask excluding non-tumoral vasculature. The refined tumor mask (after LBV removal) was subsequently applied to all parametric maps for further statistical analysis.\u003c/p\u003e\n\u003cp\u003e2.7 Statistical Feature Selection\u003c/p\u003e\n\u003cp\u003eStatistical features were extracted from the tumor regions on each parametric map (rCBV, rCBF, Slope-2, K\u003csup\u003etrans\u003c/sup\u003e, V\u003csub\u003ee\u003c/sub\u003e, and V\u003csub\u003ep\u003c/sub\u003e) after LBV removal. These features included minimum value, maximum value, mean, standard deviation, 10\u003csup\u003eth\u003c/sup\u003e, 75\u003csup\u003eth\u003c/sup\u003e and 95\u003csup\u003eth\u003c/sup\u003e percentile. Thus, for further analysis, a total of 42 features (6 parameters \u0026times; 7 statistical measures) were initially computed. Moreover, among all the slices containing tumor regions, the study was performed on 5 consecutive slices around the center of the tumor to reduce bias in the overall dataset.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.8 Statistical Analysis\u003c/p\u003e\n\u003cp\u003eThe normality of feature distributions within each glioma subtype was assessed using the Shapiro-Wilk test. For each feature, data were grouped according to class, and if all groups passed the test for normality, t-test was performed. If normality was not established in any group, the non-parametric Mann Whitney U-test was used for comparing distributions of perfusion parameters between mGBM and DMG groups. Features with p-value less than 0.05 in the respective test were considered statistically significant.\u003c/p\u003e\n\u003cp\u003eReceiver operating characteristic (ROC) curve analysis was also performed to evaluate the discriminatory performance of each individual perfusion parameter in differentiating between mGBM and DMG. For each parameter, area under the ROC curve (AUC), sensitivity, specificity, and optimal cutoff values were calculated based on Youden\u0026apos;s index, and 95% confidence intervals using the DeLong method. Statistical significance was assessed at p \u0026lt; 0.05. Box-and-whisker plots and violin plots were generated to visualize the distribution of 95\u003csup\u003eth\u003c/sup\u003e percentile values across tumor types, displaying median, mean, quartiles, and outliers.\u003c/p\u003e\n\u003cp\u003e2.9 Machine Learning Based Classification\u003c/p\u003e\n\u003cp\u003eTo evaluate multi-parametric approaches for glioma subtype differentiation, a comprehensive supervised machine learning (ML) framework was implemented using multiple classical classifiers with rigorous feature selection.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.9.1 Data Preparation and Splitting\u003c/p\u003e\n\u003cp\u003eThe dataset was split into training (70%) and testing (30%) subsets, stratified by class labels to preserve class distribution. Thus, in this study, among 62 patient\u0026rsquo;s dataset, 43 (21 mGBM, 22 DMG) were used for training and 19 (9 mGBM, 10 DMG) for testing. \u0026nbsp;Prior to model training, standardization was performed via z-score scaling to ensure equal feature importance.\u003c/p\u003e\n\u003cp\u003e2.9.2 Feature Selection Strategy\u003c/p\u003e\n\u003cp\u003eTo identify optimal feature subsets, we systematically evaluated all possible combinations of the top three selected 95\u003csup\u003eth\u003c/sup\u003e percentile perfusion parameters (rCBV, rCBF, V\u003csub\u003ee\u003c/sub\u003e). This exhaustive approach comprehensively tested single (3 combinations), two (3 combinations), three (1 combination) feature models. This resulted in 7 unique feature combinations evaluated across all four ML classifiers, yielding 28 total model configurations (7 combinations \u0026times; 4 classifiers).\u003c/p\u003e\n\u003cp\u003e2.9.3 Classification Models\u003c/p\u003e\n\u003cp\u003eFour ML algorithms were implemented for binary classification. Random forest (RF) with 100 estimators, logistic regression (LR) with L2 regularization and 500 maximum iterations using L-BFGS (Limited-memory Broyden\u0026ndash;Fletcher\u0026ndash;Goldfarb\u0026ndash;Shanno) solver, support vector machine (SVM) using radial basis function kernel with probability estimates enabled, and k-nearest neighbors (KNN) with 5 nearest neighbors, distance weights, and manhattan metrics. All models used random state 42 for reproducibility. Prior to final evaluation, hyperparameter optimization was performed using GridSearchCV with 5-fold cross-validation (CV) to identify optimal model configurations.\u003c/p\u003e\n\u003cp\u003e2.9.4 Model Evaluation \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eModel performance was comprehensively assessed using 5-fold stratified CV on the training set. The best feature subset for each model-class pair combination was selected based on maximum CV accuracy. Final model performance was evaluated on the held-out validation set using accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). Statistical significance was assessed at p \u0026lt; 0.05.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.10 Model Implementation\u003c/p\u003e\n\u003cp\u003eAll statistical and ML-based analyses were performed in Python (v3.9.12, 64-bit, Anaconda distribution) on a system with an AMD Ryzen 7 5700U processor with Radeon Graphics and 16 GB of memory. Data preprocessing was carried out using pandas (v1.4.2) and NumPy (v1.21.5). Machine learning workflows were implemented with scikit-learn (v1.0.2), while statistical analyses used SciPy (v1.7.3) and stats models (v0.13.2). Visualizations were generated with Matplotlib (v3.5.1) and Seaborn (v0.11.2).\u0026nbsp;\u003c/p\u003e"},{"header":"3 Results","content":"\u003cp\u003e3.1 Patient Characteristics\u003c/p\u003e\n\u003cp\u003eThe demographic and clinical characteristics of the study population are summarized in Table 2. DMG patients were significantly younger (mean age: 23.9 \u0026plusmn; 15.1 years) compared to mGBM (30.5 \u0026plusmn; 9.3 years) patients (p = 0.239). Gender distribution showed male predominance in mGBM (63.3% male) with a more balanced distribution in DMG (56.3% male).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMolecular profiling showed all DMGs (n=32) were H3 K27-altered while all mGBMs (n=30) cases were lacking this alteration (p \u0026lt; 0.001). Both groups were exclusively IDH-wildtype, consistent with their classification as WHO grade 4 tumors.\u003c/p\u003e\n\u003cp\u003eTumor location varied significantly between groups, with DMGs predominantly involving midline structure including thalamus (n=18, 56.3%), and brainstem (n=14, 43.7%). while mGBMs were distributed across supratentorial midline locations including thalamus (n=19, 63.3%), corpus callosum (n=7, 23.3%), and other midline structures (n=4, 13.3%) (p \u0026lt; 0.001).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u0026nbsp;\u003c/strong\u003eDemographic, clinical, and conventional MRI characteristics of the study population.\u0026nbsp;\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e\u003cstrong\u003emGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDMG\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eLocation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003eSupratentorial midline structure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003eMidline structure\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eAge (Mean \u003cstrong\u003e\u0026plusmn;\u0026nbsp;\u003c/strong\u003eSD years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e30.5 \u0026plusmn; 9.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003e23.9 \u0026plusmn; 15.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eGender (Males/Females)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e19/11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003e18/14\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eIDH status\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003eWildtype\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003eWildtype\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eH3K27-alteration\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003eAbsent\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003ePresent\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eContrast enhancement\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e29/30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003e23/32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eNecrosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e29/30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003e11/32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e3.2 Conventional MRI Features\u003c/p\u003e\n\u003cp\u003eThe conventional MRI characteristics showed significant differences between tumor types (Table 2). Contrast enhancement was observed predominantly in mGBM (96.7%) but present in only 71.8% of DMG cases (p \u0026lt; 0.001), with necrosis also most frequent in mGBM (p \u0026lt; 0.001). Figure 3 presents representative examples of mGBM and DMG cases, illustrating both enhanced variants (Case 1) showing strong gadolinium uptake, and non-enhanced variants (Case 2) demonstrating minimal or absent enhancement despite their aggressive WHO grade 4 classification. The figure displays each variant\u0026rsquo;s FLAIR, T1-Gd, rCBV, and rCBV maps after removing the LBVs. This visualization highlights the challenges of differentiating between mGBM and DMG using qualitative assessment of conventional imaging alone, emphasizing the need for quantitative perfusion analysis.\u003c/p\u003e\n\u003cp\u003e3.4 Statistical-based Classification\u003c/p\u003e\n\u003cp\u003eAmong all statistical features extracted from parametric maps, the 95\u003csup\u003eth\u003c/sup\u003e percentile values demonstrated the highest discriminatory power for differentiating between mGBM and DMG. In Figure 4, both the box and whisker plots (Figure 4a) and violin plots (Figure 4b) demonstrate clear separation between the two midline tumor types across the 95\u003csup\u003eth\u003c/sup\u003e percentile values of rCBV, rCBF, and V\u003csub\u003ee\u003c/sub\u003e. Both plots reveal a significant reduction in the median, mean, and their standard deviation values across DMG from mGBM. For rCBV, mGBM exhibited significantly higher values (mean 95\u003csup\u003eth\u003c/sup\u003e percentile: 2.11 \u0026plusmn; 1.15) compared to DMG (1.45 \u0026plusmn; 0.39). Similarly, rCBF and V\u003csub\u003ee\u003c/sub\u003e values showed a statistically significant decrease in DMG from mGBM with minimal overlap between groups, but the medians remain distinctly separated. The violin plots further confirmed that the distribution density for mGBM is skewed toward higher values, whereas DMG exhibits a more compact distribution centered at lower values.\u003c/p\u003e\n\u003cp\u003eNormality of the data distributions for each glioma subtype was assessed using the Shapiro-Wilk test. In this study, the test results indicated that the assumption of normality (p \u0026lt; 0.05) across features and classes was right and given these findings, parametric statistical methods were employed. Specifically, the t-test was applied which confirmed statistically significant differences for all perfusion parameters between mGBM and DMG (p \u0026lt; 0.05). These significant group differences are summarized in Table 3.\u003c/p\u003e\n\u003cp\u003eROC curve analysis was performed to quantitatively assess the discriminatory performance of each parameter in differentiating between mGBM and DMG. Figure 5 and Table 3 present the discriminatory performance of individual perfusion parameters for distinguishing the midline HGG subtypes. Figure 5a displays ROC curves for all six perfusion parameters and Figure 5b presents confusion matrices for the three best-performing parameters (rCBV, rCBF, and V\u003csub\u003ee\u003c/sub\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u003c/strong\u003e Statistical-based analysis performance metrics of individual perfusion parameters for distinguishing mGBM from DMG.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"626\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePerfusion Parameters\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e\u003cstrong\u003et-test\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e95% CI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOptimal Threshold\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003erCBF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e75.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003ep=0.003**\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e78.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e73.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.631-0.875\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e29.777\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003erCBV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e71.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003ep\u0026lt;0.001***\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e78.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e63.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.583-0.842\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e1.621\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003eV\u003csub\u003ee\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e70.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003ep=0.008**\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e65.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.576-0.836\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.040\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003eSlope 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e65.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003ep=0.013*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e84.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e60.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.514-0.789\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.014\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003eV\u003csub\u003ep\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e65.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003ep\u0026lt;0.013*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e87.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e53.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.513-0.787\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.018\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003eK\u003csup\u003etrans\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e55.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003ep=0.186\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e34.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e90.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.414-0.702\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 89px;\"\u003e\n \u003cp\u003e0.037\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 3 summarizes the classification metrics for each individual perfusion parameter, such as, rCBF emerged as the best single discriminator, achieving an AUC of 75.31% (95% CI: 0.631-0.875, p\u0026lt;0.001) with sensitivity of 78.12% and specificity of 73.33% at an optimal threshold of 29.777. However, K\u003csup\u003etrans\u003c/sup\u003e demonstrated the lowest discriminatory power with AUC of 55.83% (p=0.434), indicating it does not significantly distinguish between mGBM and DMG at conventional significance levels.\u003c/p\u003e\n\u003cp\u003e3.5 Machine Learning-based Classification\u003c/p\u003e\n\u003cp\u003eTo investigate the potential of combining multiple quantitative perfusion parameters, we systematically evaluated all possible feature combinations of the statistically performing top three perfusion parameters (rCBV, rCBF, V\u003csub\u003ee\u003c/sub\u003e) using four ML classifiers (RF, LR, SVM, KNN). Table 4 presents comprehensive results from this systematic analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4.\u003c/strong\u003e Machine learning based classification performance metrics for distinguishing mGBM from DMG.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"626\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSelected Features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModels\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCV Accuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" style=\"width: 321px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTest\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 3px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e(%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003erCBV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e69.72 \u0026plusmn; 5.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e68.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e66.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e72.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e74.44\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003erCBF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e72.22 \u0026plusmn; 9.17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e68.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e66.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e72.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e73.33\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003eV\u003csub\u003ee\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e74.44 \u0026plusmn; 8.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e68.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e66.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e72.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e58.89\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003erCBV + rCBF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e69.44 \u0026plusmn; 10.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e73.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e72.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e76.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e72.78\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003erCBV + V\u003csub\u003ee\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e74.72 \u0026plusmn; 7.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e57.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e57.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e66.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e65.56\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003erCBF + V\u003csub\u003ee\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e76.67 \u0026plusmn; 7.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e63.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e61.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e69.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e61.11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003erCBV + rCBF + V\u003csub\u003ee\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e76.67 \u0026plusmn; 7.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e63.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e61.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e80.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e69.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 69px;\"\u003e\n \u003cp\u003e68.89\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAbbreviations: CV, cross-validation; F1, F1-score; AUC, area under the curve; RF, random forest; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbor; rCBV, relative cerebral blood volume; rCBF, relative cerebral blood flow; V\u003csub\u003ee\u003c/sub\u003e, extravascular extracellular space volume fraction.\u003c/p\u003e\n\u003cp\u003eAmong single-parameter models, V\u003csub\u003ee\u003c/sub\u003e demonstrated the strongest cross-validation performance (RF: 74.44 \u0026plusmn; 8.90% CV accuracy), while rCBV achieved the highest test AUC (RF: 74.44%). All three single-parameter models achieved identical test accuracy of 68.42% with consistent recall of 80.00%, indicating reliable sensitivity for DMG detection.\u003c/p\u003e\n\u003cp\u003eTwo-parameter combinations demonstrated improved test accuracy compared to single parameters. The rCBV and V\u003csub\u003ee\u003c/sub\u003e combination achieved the highest CV accuracy (RF: 76.67% \u0026plusmn; 7.16%) among two-parameter models but moderate test accuracy (63.16%). Whereas the combination of rCBV and rCBF achieved the highest overall test accuracy of 73.68% using KNN, with precision of 72.73% and F1-score of 76.19%, though CV accuracy was 69.44% \u0026plusmn; 10.54%.\u003c/p\u003e\n\u003cp\u003eThe three-parameter combination (rCBV + rCBF + V\u003csub\u003ee\u003c/sub\u003e) using RF achieved the highest CV accuracy overall (76.67% \u0026plusmn; 7.16%) with relatively low standard deviation, indicating stable CV performance with improved AUC (68.89%) compared to some two-parameter combinations.\u0026nbsp;\u003c/p\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eThis study comprehensively evaluated both individual DCE-MRI perfusion parameters and ML-based multi-parametric approaches for differentiating DMG from mGBM in children and young adults under 40 years of age. Our results demonstrate that individual parameters, particularly rCBF, rCBV, and V\u003csub\u003ee\u003c/sub\u003e, achieved robust differentiation (Table 3). ML classifiers, through systematic evaluation of all possible feature combinations from these three perfusion parameters, demonstrated that carefully selected combinations can provide clinically useful discrimination. The distinct perfusion profiles observed, with mGBM exhibiting the highest parametric values compared to DMG, reflect the underlying molecular and pathophysiological characteristics of these tumor entities [4,7].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe molecular basis for these perfusion differences is well-established in the literature. mGBM is characterized by EGFR amplification, TERTp mutations, and the upregulation of vascular endothelial growth factor (VEGF), promoting robust angiogenesis and vascular permeability, [4,7,29]. In contrast, H3K27-altered DMG exhibits distinct epigenetic dysregulation affecting cellular differentiation and vascular development, resulting in less aggressive angiogenic activity [4]. The significantly lower perfusion values in DMG compared to mGBM quantitatively validate this molecular distinction at the imaging level [17,18].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe clinical imperative for non-invasive differentiation is emphasized by the inherent challenges of tissue diagnosis in these tumors. While immunohistochemistry for H3 K27M mutation remains the gold standard for definitive DMG diagnosis [2,4], obtaining adequate tissue from eloquent midline structures presents formidable obstacles. Stereotactic biopsy of thalamic and brainstem lesions carries substantial risks including hemorrhage, neurological deficits, and procedure-related mortality, with complication rates reaching 10-15% in some series [12,13]. In patients with critical tumor locations or significant medical comorbidities, surgical sampling may be absolutely contraindicated. Thus, differentiation from other midline HGG tumors remains challenging yet clinically crucial (Figure 3) [12,14,17,18].\u003c/p\u003e\n\u003cp\u003eBeyond these molecular and clinical challenges, understanding the demographic context of these tumors is essential for appreciating the significance of our findings (Table 2). DMG with H3 K27-alteration has a bimodal age distribution, affecting primarily children/adolescents and young adults. Previous studies on DMGs have often focused on pediatric populations or included very wide age ranges [3,12,13]. However, our study extends previous work by specifically focusing on DMG differentiation in a carefully selected pediatric and young adult population (age \u0026lt; 40 years), addressing an important gap in the literature. By targeting this relatively homogeneous age group, we minimized confounding effects of age-related vascular changes while maintaining clinical relevance. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eGiven the distinct characteristics of these tumors, accurate differentiation among midline HGG subtypes has important therapeutic implications for treatment planning, prognostication, and clinical trial eligibility [3,5,10]. Current treatment strategies differ substantially between these entities. mGBM typically receive maximal safe resection followed by chemoradiation [42,43], while DMGs may require alternative therapeutic approaches including targeted therapies, and clinical trial enrollment [8,11]. This therapeutic divergence underscores the importance of accurate preoperative characterization [4,5].\u003c/p\u003e\n\u003cp\u003eConventional MRI remains the foundation of brain tumor imaging but has well-recognized limitations in distinguishing between DMG and other HGG subtypes as all typically appear as heterogeneously enhancing masses with surrounding edema and variable necrosis (Figure 3) [16,17,34]. While some features may suggest specific diagnoses, such as the strict midline location of DMGs, there is a substantial overlap in their imaging appearances, particularly between mGBMs and DMGs [17,18]. Our study indicates that advanced perfusion parameters can complement conventional imaging to improve diagnostic accuracy.\u003c/p\u003e\n\u003cp\u003eQuantitative perfusion parameters effectively address this diagnostic challenge. Our findings show that both individual parameters and ML classifiers provide significant discriminatory power for this critical distinction. Specifically, rCBF, rCBV, and V\u003csub\u003ee\u003c/sub\u003e achieved AUC values of 75.31%, 71.25%, and 70.62%, respectively (Table 3), while ML classifiers achieved the highest CV accuracy of 76.67% (RF with rCBF+V\u003csub\u003ee\u003c/sub\u003e, and RF with rCBV+rCBF+V\u003csub\u003ee\u003c/sub\u003e) and highest test accuracy of 73.68% (KNN with rCBV+rCBF) (Table 4).\u003c/p\u003e\n\u003cp\u003eFrom a clinical implementation perspective, the findings have several important implications. First, quantitative DCE-MRI parameters can complement conventional MRI to improve diagnostic accuracy when molecular testing is unavailable or when biopsy is contraindicated due to eloquent tumor locations\u0026nbsp;and associated procedural risks. Second, the automated analysis pipeline including LBV segmentation, feature extraction, and ML classification could potentially be integrated into clinical workflows to provide objective, reproducible tumor characterization. Third, the identification of key discriminatory features (rCBF, rCBV, V\u003csub\u003ee\u003c/sub\u003e) suggests that simplified imaging protocols focusing on these parameters might be sufficient for clinical decision-making, potentially reducing computational burden.\u0026nbsp;Finally, this approach ensures minimal risk of missing DMG cases that might benefit from targeted therapies or eligibility for clinical trial enrollment.\u003c/p\u003e\n\u003cp\u003e4.1 Limitations\u003c/p\u003e\n\u003cp\u003eThis study also has some limitations. First, the retrospective single-institution design may limit generalizability. Second, despite covering a six-year period (2018-2024), the sample size remains modest, particularly for the test set (19 cases: 9 mGBM, 10 DMG), which may affect the reliability of ML model performance estimates. Third, DMG patients are relatively rare and histopathological confirmation was challenging due to eloquent tumor locations. Fourth, our analysis incorporated data classified under both the 2016 and 2021 WHO classification systems, which might introduce classification inconsistencies [1,2]. Fifth, we focused exclusively on DCE-MRI perfusion parameters and did not integrate other advanced imaging modalities (diffusion-weighted imaging, susceptibility imaging, MR spectroscopy) or radiomics features, which might further improve classification accuracy.\u003c/p\u003e\n\u003cp\u003e4.2 Future Directions\u003c/p\u003e\n\u003cp\u003eFuture research should focus on validation of results of this study on large multicenter data. Integration of perfusion metrics with other advanced imaging features (diffusion parameters, spectroscopic metabolites, radiomics features) in comprehensive multimodal ML models may provide superior diagnostic performance. Deep learning approaches that can automatically extract relevant features from raw parametric maps without manual feature engineering represent a promising direction.\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eThis study evaluated the potential of quantitative DCE-MRI perfusion metrics and ML classifiers for the non-invasive differentiation of DMG from mGBM in children and young adults. DMG demonstrated significantly lower perfusion values compared to mGBM, consistent with their distinct biological profiles. Individual perfusion parameters, particularly rCBF, rCBV and V\u003csub\u003ee\u0026nbsp;\u003c/sub\u003eachieved clinically useful discrimination and systematic evaluation of multi-parametric ML approaches, particularly RF provided incremental improvements. The interpretation of perfusion imaging in this study carries important implications for improving diagnostic accuracy in neuro-oncology, particularly when molecular testing through tissue acquisition is unavailable or contraindicated due to procedural risks.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eGBM, glioblastoma; mGBM, midline glioblastoma; DMG, diffuse midline glioma; HGG, high-grade glioma; MRI, magnetic resonance imaging; T1-W, T1-weighted; T2-W, T2-weighted; PD-W, proton density-weighted; FLAIR, fluid-attenuated inversion recovery; T1-Gd, postcontrast T1-W; DCE-MRI, dynamic contrast-enhanced magnetic resonance imaging; TR, repetition time; TE, echo time; FOV, field of view; PL, piecewise linear; ETM, extended Toft\u0026apos;s model; rCBV, relative cerebral blood volume; rCBF, relative cerebral blood flow; Slope-2, washout slope; K\u003csup\u003etrans\u003c/sup\u003e, volume transfer constant; V\u003csub\u003ee\u003c/sub\u003e, extravascular extracellular space volume fraction; V\u003csub\u003ep\u003c/sub\u003e, blood plasma volume fraction; AI, artificial intelligence; LBV, large blood vessel; ROC, receiver operating characteristic; ML, machine learning; RF, random forest; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbors; CV, cross validation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eConceptualization: A.K., R.K.G., A.S.; Data curation: A.K.; Formal analysis: A.K.; Investigation: A.K.; Methodology: A.K., R.K.G., A.S.; Validation: A.K., R.K.G., A.S., S.A., R.P., S.V.; Visualization: A.K.; Writing \u0026ndash; original draft: A.K.; Writing \u0026ndash; review and editing: A.K., R.K.G., A.S., S.A., R.P., S.V.; Resources: R.K.G., S.A., R.P., S.V.; Supervision R.K.G., A.S.; Funding acquisition: A.S.; Project administration: A.S.\u003c/p\u003e\n\u003ch2\u003eAcknowledgement\u003c/h2\u003e\n\u003cp\u003eThe authors thank Mr. Rakesh Kumar Singh from Fortis Memorial Research Institute, Gurugram, India for data acquisition and the MedImg Lab members for their constant guidance and support.\u003c/p\u003e\n\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eThe code and a sample of anonymized test data can be accessed after sending an appropriate email request to the corresponding author.\u003c/p\u003e\n\u003ch2\u003eEthics and Consent to Participate Declarations\u003c/h2\u003e\n\u003cp\u003eThis study was approved by the institutional ethics committee of XYZ Hospital (name anonymized per journal requirements), and the requirement for informed consent was waived by the committee.\u003c/p\u003e\n\u003ch2\u003eFunding Declaration\u003c/h2\u003e\n\u003cp\u003eThis work was supported by the Science and Engineering Research Board, Department of Science \u0026amp; Technology (project number: CRG/2019/005032) and Indian Council of Medical Research, Government of India (Project number: CAR-2024-01-000187).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eLouis DN, Perry A, Reifenberger G, von Deimling A, Figarella-Branger D, Cavenee WK, Ohgaki H, Wiestler OD, Kleihues P, Ellison DW (2016) The 2016 World Health Organization Classification of tumors of the central nervous system: a summary. Acta Neuropathol 131:803\u0026ndash;820. https://doi.org/10.1007/s00401-016-1545-1\u003c/li\u003e\n\u003cli\u003eLouis DN, Perry A, Wesseling P, Brat DJ, Cree IA, Figarella-Branger D, Hawkins C, Ng HK, Pfister SM, Reifenberger G, Soffietti R, von Deimling A, Ellison DW (2021) The 2021 WHO classification of tumors of the central nervous system: a summary. Neuro Oncol 23:1231\u0026ndash;1251. https://doi.org/10.1093/neuonc/noab106\u003c/li\u003e\n\u003cli\u003eZheng L, Gong J, Yu T, Zou Y, Zhang M, Nie L, Chen X, Yue Q, Liu Y, Mao Q, Zhou Q, Chen N (2022) Diffuse midline gliomas with histone H3 K27M mutation in adults and children: a retrospective series of 164 cases. Am J Surg Pathol 46:863\u0026ndash;871. https://doi.org/10.1097/PAS.0000000000001897\u003c/li\u003e\n\u003cli\u003ePark YW, Vollmuth P, Foltyn-Dumitru M, Sahm F, Ahn SS, Chang JH, Kim SH (2023) The 2021 WHO classification for gliomas and implications on imaging diagnosis: Part 2\u0026mdash;summary of imaging findings on pediatric-type diffuse high-grade gliomas, pediatric-type diffuse low-grade gliomas, and circumscribed astrocytic gliomas. J Magn Reson Imaging 58:690\u0026ndash;708. https://doi.org/10.1002/jmri.28543\u003c/li\u003e\n\u003cli\u003eAl Sharie S, Abu Laban D, Al-Hussaini M (2023) Decoding diffuse midline gliomas: a comprehensive review of pathogenesis, diagnosis and treatment. Cancers (Basel) 15:4869. https://doi.org/10.3390/cancers15194869\u003c/li\u003e\n\u003cli\u003eLovibond S, Gewirtz AN, Pasquini L, Krebs S, Graham MS (2023) The promise of metabolic imaging in diffuse midline glioma. Neoplasia 39:100896. https://doi.org/10.1016/j.neo.2023.100896\u003c/li\u003e\n\u003cli\u003ePark YW, Vollmuth P, Foltyn-Dumitru M, Sahm F, Ahn SS, Chang JH, Kim SH (2023) The 2021 WHO classification for gliomas and implications on imaging diagnosis: Part 1\u0026mdash;key points of the fifth edition and summary of imaging findings on adult-type diffuse gliomas. J Magn Reson Imaging 58:677\u0026ndash;689. https://doi.org/10.1002/jmri.28743 \u003c/li\u003e\n\u003cli\u003eAkdemir EY, Odia Y, Hall MD, Mehta MP, Kotecha R (2024) An update on H3K27M-altered diffuse midline glioma: diagnostic and therapeutic challenges in clinical practice. Pract Radiat Oncol 14:443\u0026ndash;451. https://doi.org/10.1016/j.prro.2024.04.013\u003c/li\u003e\n\u003cli\u003eJoshi S, Wu YK, Xu YK, Liu XM, Zhang H (2022) Comparative study of diffuse midline glioma and glioblastoma: Magnetic Resonance Imaging (MRI) in the characteristics and demography. Nepal J Neurosci 19:32\u0026ndash;36. https://doi.org/10.3126/njn.v19i2.43724\u003c/li\u003e\n\u003cli\u003ePachocki CJ, Hol EM (2022) Current perspectives on diffuse midline glioma and a different role for the immune microenvironment compared to glioblastoma. J Neuroinflammation 19:212. https://doi.org/10.1186/s12974-022-02630-8\u003c/li\u003e\n\u003cli\u003eChi AS, Tarapore RS, Hall MD, et al (2019) Pediatric and adult H3 K27M-mutant diffuse midline glioma treated with the selective DRD2 antagonist ONC201. J Neurooncol 145:97\u0026ndash;105. https://doi.org/10.1007/s11060-019-03271-3\u003c/li\u003e\n\u003cli\u003eSim Y, Choi SH, Lee N, Park YW, Ahn SS, Chang JH, Kim SH, Lee SK (2024Clinical, qualitative imaging biomarkers, and tumor oxygenation imaging biomarkers for differentiation of midline-located IDH wild-type glioblastomas and H3 K27-altered diffuse midline gliomas in adults. Eur J Radiol 173:111384. https://doi.org/10.1016/j.ejrad.2024.111384\u003c/li\u003e\n\u003cli\u003eWiśniewski K, Ghaly A, Drummond K, Fahlstr\u0026ouml;m A (2023) H3 K27M-altered diffuse midline gliomas: a review. Indian J Neurosurg 12:104\u0026ndash;115. https://doi.org/10.1055/s-0043-1771192\u003c/li\u003e\n\u003cli\u003eLv K, Chen H, Cao X, Du P, Chen J, Liu X, Zhu L, Geng D, Zhang J (2023) Development and validation of a machine learning algorithm for predicting diffuse midline glioma, H3 K27\u0026ndash;altered, H3 K27 wild-type high-grade glioma, and primary CNS lymphoma of the brain midline in adults. J Neurosurg 39:393\u0026ndash;401. https://doi.org/10.3171/2022.11.JNS221544\u003c/li\u003e\n\u003cli\u003eChauhan RS, Kulanthaivelu K, Kathrani N, Kotwal A, Bhat MD, Saini J, Srinivas D (2021) Prediction of H3K27M mutation status of diffuse midline gliomas using MRI features. J Neuroimaging 31:1201\u0026ndash;1210. https://doi.org/10.1111/jon.12905\u003c/li\u003e\n\u003cli\u003eChauhan RS, Kulanthaivelu K, Kathrani N, et al (2021) Can conventional MRI features predict H3K27M mutation status of diffuse midline gliomas? Res Square Preprint. https://doi.org/10.21203/rs.3.rs-360529/v1\u003c/li\u003e\n\u003cli\u003eBanan R, Akbarian A, Samii M, Samii A, Bertalanffy H, Lehmann U, Hartmann C, Br\u0026uuml;ning R (2021) Diffuse midline gliomas, H3 K27M-mutant are associated with less peritumoral edema and contrast enhancement in comparison to glioblastomas, H3 K27M-wildtype of midline structures. PLoS One 16:e0249647. https://doi.org/10.1371/journal.pone.0249647\u003c/li\u003e\n\u003cli\u003eRaab P, Banan R, Akbarian A, et al (2022) Differences in the MRI Signature and ADC Values of Diffuse Midline Gliomas with H3 K27M Mutation Compared to Midline Glioblastomas. Cancers (Basel) 14:1397. https://doi.org/10.3390/cancers14061397\u003c/li\u003e\n\u003cli\u003eKim M, Jung SY, Park JE, Jo Y, Park SY, Nam SJ, Kim JH, Kim HS (2020) Diffusion- and perfusion-weighted MRI radiomics model may predict isocitrate dehydrogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade glioma. Eur Radiol 30:2142\u0026ndash;2151. https://doi.org/10.1007/s00330-019-06548-3\u003c/li\u003e\n\u003cli\u003eGuo W, She D, Xing Z, Lin X, Wang F, Song Y, Cao D (2022) Multiparametric MRI-Based Radiomics Model for Predicting H3 K27M Mutant Status in Diffuse Midline Glioma: A Comparative Study Across Different Sequences and Machine Learning Techniques, Front Oncol 12:796583. https://doi.org/10.3389/FONC.2022.796583\u003c/li\u003e\n\u003cli\u003eKurokawa R, Baba A, Kurokawa M, et al (2022) Perfusion and diffusion-weighted imaging parameters: Comparison between pre-and postbiopsy MRI for high-grade glioma. Medicine 101:e30183. https://doi.org/10.1097/MD.0000000000030183\u003c/li\u003e\n\u003cli\u003eRameh V, Vajapeyam S, Ziaei A, et al (2023) Correlation between Multiparametric MR Imaging and Molecular Genetics in Pontine Pediatric High-Grade Glioma. AJNR Am J Neuroradiol 44:833\u0026ndash;840. https://doi.org/10.3174/ajnr.A7910\u003c/li\u003e\n\u003cli\u003eSzychot E, Youssef A, Ganeshan B, et al (2021) Predicting outcome in childhood diffuse midline gliomas using magnetic resonance imaging based texture analysis. J Neuroradiol 48:243\u0026ndash;247. https://doi.org/10.1016/j.neurad.2020.02.005\u003c/li\u003e\n\u003cli\u003eKathrani N, Chauhan RS, Kotwal A, et al (2022) Diffusion and perfusion imaging biomarkers of H3 K27M mutation status in diffuse midline gliomas. Neuroradiology 64:1519\u0026ndash;1528. https://doi.org/10.1007/s00234-021-02857-x\u003c/li\u003e\n\u003cli\u003eWu C, Zheng H, Li J, et al (2022) MRI-based radiomics signature and clinical factor for predicting H3K27M mutation in pediatric high-grade gliomas located in the midline of the brain. Eur Radiol 32:1813\u0026ndash;1822. https://doi.org/10.1007/s00330-021-08234-9\u003c/li\u003e\n\u003cli\u003eKurokawa R, Kurokawa M, Baba A, et al (2023) Dynamic susceptibility contrast perfusion-weighted and diffusion-weighted magnetic resonance imaging findings in pilocytic astrocytoma and H3.3 and H3.1 variant diffuse midline glioma, H3K27-altered. PLoS One 18:e0288412. https://doi.org/10.1371/journal.pone.0288412\u003c/li\u003e\n\u003cli\u003eCramer S, Popa L, Haley S, et al (2020) Path-04. the blood-brain barrier in diffuse midline glioma and its implications for drug delivery. Neuro Oncol 22:ii164. https://doi.org/10.1093/neuonc/noaa215.686\u003c/li\u003e\n\u003cli\u003eTofts PS, Kermode AG (1991) measurement of the blood-brain barrier permeability and leakage space using dynamic MR imaging. 1. fundamental concepts. Magn Reson Med 17:357\u0026ndash;367. https://doi.org/10.1002/mrm.1910170223\u003c/li\u003e\n\u003cli\u003eCuenod CA, Balvay D (2013) Perfusion and vascular permeability: basic concepts and measurement in DCE-CT and DCE-MRI. Diagn Interv Imaging 94:1187\u0026ndash;1204. https://doi.org/10.1016/j.diii.2013.10.010\u003c/li\u003e\n\u003cli\u003eJain R, Griffith B, Alotaibi F, et al (2015) Glioma angiogenesis and perfusion imaging: Understanding the relationship between tumor blood volume and leakiness with increasing glioma grade. AJNR Am J Neuroradiol 36:2030\u0026ndash;2035. https://doi.org/10.3174/ajnr.A4405\u003c/li\u003e\n\u003cli\u003eChen YJ, Chu WC, Pu YS, et al (2012) Washout gradient in dynamic contrast-enhanced MRI is associated with tumor aggressiveness of prostate cancer. J Magn Reson Imaging 36:912\u0026ndash;919. https://doi.org/10.1002/jmri.23723\u003c/li\u003e\n\u003cli\u003eGupta PK, Saini J, Sahoo P, et al (2017) Role of Dynamic Contrast-Enhanced Perfusion Magnetic Resonance Imaging in Grading of Pediatric Brain Tumors on 3T. Pediatr Neurosurg 52:298\u0026ndash;305. https://doi.org/10.1159/000479283\u003c/li\u003e\n\u003cli\u003eLaw M, Yang S, Babb JS, et al (2004) Comparison of Cerebral Blood Volume and Vascular Permeability from Dynamic Susceptibility Contrast-Enhanced Perfusion MR Imaging with Glioma Grade. AJNR Am J Neuroradiol 25:746\u0026ndash;755\u003c/li\u003e\n\u003cli\u003eSaini J, Gupta PK, Sahoo P, et al (2018) Differentiation of grade II/III and grade IV glioma by combining \u0026ldquo;T1 contrast-enhanced brain perfusion imaging\u0026rdquo; and susceptibility-weighted quantitative imaging. Neuroradiology 60:43\u0026ndash;50. https://doi.org/10.1007/s00234-017-1942-8 \u003c/li\u003e\n\u003cli\u003eSengupta A, Agarwal S, Gupta PK, et al (2018) On differentiation between vasogenic edema and non-enhancing tumor in high-grade glioma patients using a support vector machine classifier based upon pre and post-surgery MRI images. Eur J Radiol 106:199\u0026ndash;208. https://doi.org/10.1016/j.ejrad.2018.07.018\u003c/li\u003e\n\u003cli\u003eSingh A, Rathore RKS, Haris M, et al (2009) Improved bolus arrival time and arterial input function estimation for tracer kinetic analysis in DCE-MRI. J Magn Reson Imaging 29:166\u0026ndash;176. https://doi.org/10.1002/jmri.21624\u003c/li\u003e\n\u003cli\u003eSingh A, Haris M, Rathore D, et al (2007) Quantification of physiological and hemodynamic indices using T1 dynamic contrast- enhanced MRI in intracranial mass lesions. J Magn Reson Imaging 26:871\u0026ndash;880. https://doi.org/10.1002/jmri.21080 \u003c/li\u003e\n\u003cli\u003eAshburner J, Barnes G, Chen CC, et al (2021) SPM12 Manual The FIL Methods Group (and honorary members), London. https://www.fil.ion.ucl.ac.uk/spm/\u003c/li\u003e\n\u003cli\u003eSahoo P, Gupta RK, Gupta PK, et al (2017 Diagnostic accuracy of automatic normalization of CBV in glioma grading using T1- weighted DCE-MRI. Magn Reson Imaging 44:32\u0026ndash;37. https://doi.org/10.1016/j.mri.2017.08.003\u003c/li\u003e\n\u003cli\u003eMaurya S, Yadav VK, Agarwal S, Singh A (2021) Brain Tumor Segmentation in mpMRI scans (BraTS-2021) using models based on U-Net Architecture, In: International MICCAI Brainlesion Workshop. LNCS 12963:312\u0026ndash;323. https://doi.org/10.1007/978-3-031-09002-8_28\u003c/li\u003e\n\u003cli\u003eKesari A, Maurya S, Sheikh MT, Gupta RK, Singh A (2025) Large blood vessel segmentation in quantitative DCE-MRI of brain tumors: A Swin UNETR approach. Magn Reson Imaging 118:110342. https://doi.org/10.1016/j.mri.2025.110342\u003c/li\u003e\n\u003cli\u003eMajewska P, Ioannidis S, Raza MH, et al (2017) Postprogression survival in patients with glioblastoma treated with concurrent chemoradiotherapy: a routine care cohort study. CNS Oncol 6:307. https://doi.org/10.2217/cns-2017-0001\u003c/li\u003e\n\u003cli\u003eStupp R, Mason WP, van den Bent MJ, et al (2005) Radiotherapy plus Concomitant and Adjuvant Temozolomide for Glioblastoma. N Engl J Med 352:987\u0026ndash;996. https://doi.org/10.1056/NEJMoa043330\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Midline glioblastoma, Diffuse midline glioma, Dynamic contrast-enhanced MRI, Perfusion parameters, Machine learning, Sequential forward selection","lastPublishedDoi":"10.21203/rs.3.rs-8481592/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8481592/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e \u003cp\u003eDiffuse midline glioma (DMG) and midline glioblastoma (mGBM) are aggressive WHO grade 4 tumors with comparable median survival of 12\u0026ndash;18 months but require fundamentally different therapeutic approaches. Despite their clinical urgency, non-invasive differentiation remains challenging due to overlapping conventional MRI features and the difficulty of obtaining tissue diagnosis from eloquent midline locations.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis retrospective study included 62 patients with histologically confirmed midline gliomas (30 mGBM, 32 DMG) evaluated with 3T MRI. Quantitative DCE-MRI perfusion parameters (rCBV, rCBF, Slope-2, K\u003csup\u003etrans\u003c/sup\u003e, V\u003csub\u003ep\u003c/sub\u003e, V\u003csub\u003ee\u003c/sub\u003e) were computed and compared between the midline tumor types. Statistical analyses included Shapiro-Wilk test, t-test, and ROC curve analysis using perfusion parameters. Machine learning-based classification was also performed using four classifiers and 5-fold cross-validation, evaluating all possible feature combinations among the best features from the perfusion parameters.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe 95th percentile values of perfusion parameters demonstrated superior discriminative capability between mGBM and DMG. DMG exhibited significantly lower perfusion parameter values compared to mGBM (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Individual perfusion parameters, particularly rCBF, rCBV, V\u003csub\u003ee\u003c/sub\u003e showed discriminative performance achieving AUC values ranging from 70.62% to 75.31%, for differentiating mGBM vs DMG. Machine learning classifiers used these features for evaluating 7 combinations. Three parameter combination (rCBV\u0026thinsp;+\u0026thinsp;rCBF\u0026thinsp;+\u0026thinsp;V\u003csub\u003ee\u003c/sub\u003e) using RF achieved highest cross-validation accuracy (76.67\u0026thinsp;\u0026plusmn;\u0026thinsp;7.16%) with consistent sensitivity (80.00%) across all models.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eQuantitative DCE-MRI perfusion analysis provides significant diagnostic value for differentiating DMG from mGBM, offering a non-invasive alternative when tissue diagnosis is not obtained. Both individual parameters and optimized multi-parametric approaches demonstrate clinically useful performance for guiding treatment decisions.\u003c/p\u003e","manuscriptTitle":"Non-Invasive Differentiation of Diffuse Midline Glioma and Midline Glioblastoma Using DCE-MRI Perfusion Parameters and Machine Learning Classification in Pediatric and Young Adult Patients","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-16 09:02:52","doi":"10.21203/rs.3.rs-8481592/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ea923bbb-1c7a-4235-919e-d18cf39f2b43","owner":[],"postedDate":"January 16th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-01-16T09:02:52+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-16 09:02:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8481592","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8481592","identity":"rs-8481592","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}