Skip to main content
NCBI home page
Frontiers in Oncology logoLink to Frontiers in Oncology
. 2026 Jul 20;16:1892760. doi: 10.3389/fonc.2026.1892760

Radiomics and clinical data predict pseudoprogression after radiotherapy in high-grade glioma

Jiang Zhou 1,2,†,‡, Zhang Danmeng 1,2,†,‡, Yang Hui 1,2,‡, Xu Zhuohua 1,2,‡, Wei Mingjing 1,2,‡, Lu Ying 1,2,*,‡
PMCID: PMC13429429  PMID: 42548620

Abstract

Background

On conventional magnetic resonance imaging, pseudoprogression after radiotherapy for high-grade glioma may closely resemble true tumor progression, leading to unnecessary surgery, premature treatment escalation, or delayed appropriate therapy. We developed and internally validated an exploratory multivariable model for individualized pseudoprogression risk estimation.

Methods

This single-center retrospective cohort included 222 patients with World Health Organization 2021 central nervous system grade 3 or 4 glioma who underwent surgery followed by radiotherapy at Liuzhou Workers’ Hospital from January 2015 to December 2024. Pseudoprogression was adjudicated before modeling through multidisciplinary review; 13 pseudoprogression cases had histopathological confirmation, and non-histologically confirmed cases required at least 6 months of stable or improved follow-up imaging. The complete two-reader radiomics matrix was analyzed using a strict split-first workflow. Dataset partitioning preceded ICC filtering, Z-score normalization, and LASSO modeling.

Results

A total of 3,404 radiomic features were analyzed. After reproducibility filtering and LASSO selection, 17 radiomic features were retained to construct the locked RadScore. In the held-out validation cohort, the integrated model combining RadScore, rCBV, ADC, NLR, MGMT promoter methylation, and TMZ treatment achieved an AUC of 0.811 (95% CI: 0.696-0.925). The clinical-imaging model without RadScore achieved a validation AUC of 0.744 (95% CI: 0.614-0.873), whereas RadScore alone achieved an AUC of 0.771 (95% CI: 0.648-0.894).

Conclusion

The integrated model showed acceptable internal validation performance and provided a clinically interpretable framework for individualized PsP risk estimation by combining radiomic, perfusion-diffusion, inflammatory, molecular, and treatment-related information. External validation with standardized imaging protocols is warranted.

Keywords: high-grade glioma, nomogram, prediction model, pseudoprogression, radiomics

Introduction

High-grade gliomas (HGGs) are the most common primary malignant tumors of the central nervous system and remain clinically aggressive, with marked invasiveness, biological heterogeneity, and poor prognosis (1). Standard treatment usually consists of maximal safe resection followed by radiotherapy with temozolomide (TMZ); however, survival remains limited despite this multimodal approach (2–4). After radiotherapy, treatment-related pseudoprogression (PsP) can be difficult to separate from true progression (TP) on conventional magnetic resonance imaging (MRI) (5).

PsP is generally regarded as a treatment-related imaging phenomenon driven by radiation-induced inflammation, increased vascular permeability, blood-brain barrier disruption, and focal necrosis rather than viable tumor growth. It may appear as new or enlarging enhancement within the irradiation field, often accompanied by edema, and is most often encountered during the first several months after radiotherapy. Reported incidence varies widely because studies differ in diagnostic criteria, treatment regimens, molecular subgroups, and follow-up duration (6–9). Misclassifying PsP as TP may lead to unnecessary surgery, inappropriate treatment escalation, or premature discontinuation of effective therapy; conversely, misclassifying TP as PsP may delay needed treatment (10).

Current assessment still relies heavily on Response Assessment in Neuro-Oncology (RANO)-based longitudinal follow-up, which requires serial imaging and can be influenced by reader judgment (11, 12). Diffusion-weighted and perfusion-weighted imaging provide additional biological information, but performance varies with acquisition protocols, post-processing, lesion timing, and region-of-interest definition (13, 14). Radiomics offers a complementary quantitative strategy by extracting high-dimensional descriptors of intensity, shape, and texture that are not fully captured by visual interpretation (15–18). Previous radiomics studies have reported encouraging discrimination between PsP and TP, yet imaging-only models may generalize poorly and may not capture the clinical, molecular, inflammatory, and treatment-related mechanisms underlying PsP (19–23).

In this context, we aimed to develop and internally validate a diagnostic multivariable model integrating MRI radiomics with clinical, molecular, treatment-related, diffusion, and perfusion variables to estimate the probability of PsP after radiotherapy in patients with HGG. Model performance was evaluated in terms of discrimination, calibration, and clinical utility using receiver operating characteristic analysis, calibration assessment, split-sample internal validation, and decision curve analysis. This article is reported in accordance with the TRIPOD checklist.

Materials and methods

Study design and participants

This retrospective prediction-model development and internal validation study was based on a consecutive cohort from Liuzhou Workers’ Hospital, a single tertiary center. We screened patients with pathologically confirmed World Health Organization 2021 central nervous system grade 3 or 4 glioma who underwent surgery followed by postoperative radiotherapy between January 2015 and December 2024. Because original pathology reports had been issued across different calendar years, all diagnoses and WHO grade assignments were re-reviewed according to the 2021 WHO Classification of Tumours of the Central Nervous System before analysis. The intended clinical application was risk estimation at the time of first suspected radiological progression after radiotherapy, when clinicians must decide whether new or enlarging enhancement is more likely to represent PsP or TP. Outcome follow-up was based on the availability of 1-year post-radiotherapy MRI or the last eligible MRI within the follow-up period.

Eligible patients had pathologically confirmed HGG, received postoperative radiotherapy, had postoperative MRI available, including T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and T2-fluid-attenuated inversion recovery sequences, and had adequate follow-up imaging for PsP assessment. Patients were excluded if clinical, imaging, molecular, or outcome data required for model development were incomplete; if image quality was insufficient for segmentation or feature extraction; or if antitumor treatment had been administered before the index surgery or radiotherapy course.

Sex was abstracted from medical records and reported as female or male according to recorded biological sex. Race and ethnicity were not included because they were not routinely recorded in the source database and were not prespecified as clinically relevant for this single-center cohort. Predictor and outcome missingness was handled by complete-case analysis; no imputation was performed.

The study protocol, entitled “Construction of a Multimodal Prediction Model for Pseudoprogression After Radiotherapy in High-Grade Glioma”, was approved by the Medical Research Ethics Committee of Liuzhou Workers’ Hospital (approval No. KY2026255). The study was conducted in accordance with the Declaration of Helsinki and subsequent amendments. The requirement for individual informed consent was waived because of the retrospective design.

Outcome definition

The target outcome was PsP after radiotherapy. PsP was diagnosed according to RANO/RANO 2.0-based longitudinal criteria: new or increased enhancement within the irradiation field after radiotherapy, no clear clinical deterioration attributable to tumor progression, and subsequent stability or improvement on follow-up imaging without treatment change. Thirteen PsP cases underwent repeat surgery or biopsy with histopathological confirmation. For patients without histopathological confirmation, a minimum of 6 months of stable or improved follow-up imaging after the suspicious progression event was required before classification as PsP. Outcome adjudication was performed through multidisciplinary review by two radiologists and two medical oncologists, and radiomic features were not used for outcome assignment. Cases with unresolved clinical-radiological evolution were excluded before formation of the final 222-patient analytic cohort; 21 indeterminate cases were excluded for this reason.

Candidate predictors

Candidate predictors were prespecified on the basis of clinical relevance and prior literature. Collected variables included sex, age, Eastern Cooperative Oncology Group performance status, extent of resection, radiotherapy dose and technique, tumor volume, edema, tumor location, seizure at presentation, isocitrate dehydrogenase 1 (IDH1) mutation status, O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status, TMZ treatment, neutrophil-to-lymphocyte ratio (NLR), relative cerebral blood volume (rCBV), apparent diffusion coefficient (ADC), dexamethasone dose, and radiomics score (RadScore). MRI radiomic features, ADC, and rCBV were derived from the assessment performed at first suspected radiological progression after radiotherapy. NLR was calculated from peripheral neutrophil and lymphocyte counts obtained within 1 week before or after the corresponding MRI assessment. TMZ treatment was coded as positive if the patient received TMZ at any time during postoperative oncologic treatment. Dexamethasone dose was recorded as the cumulative dose during treatment. Molecular markers were obtained from postoperative pathology reports after 2021 WHO re-review.

Molecular marker assessment: Original pathology reports were issued across different years and were subsequently re-reviewed according to the 2021 WHO Classification of Tumors of the Central Nervous System before analysis. IDH1-R132H mutation status was assessed by immunohistochemistry (IHC). Positive IHC staining was defined as unequivocal brown-yellow cytoplasmic and/or nuclear staining in tumor cells; cases with staining in at least 10% of tumor cells were classified as IDH1-R132H mutant, whereas scattered staining confined to non-neoplastic cells such as infiltrating macrophages was considered negative. MGMT promoter methylation status was assessed using qualitative methylation-specific polymerase chain reaction (MSP). Cases were classified as MGMT methylation positive when the methylated-primer amplification band (M band) was clearly visible, or when its optical density or abundance was greater than or equal to that of the unmethylated band (U band).

MRI acquisition and preprocessing

Radiomics analysis used postoperative follow-up MRI at the first suspected radiological progression after radiotherapy. MRI examinations were performed on two 3.0-T scanner platforms: a Siemens Healthineers MAGNETOM Skyra system equipped with a 32-channel head/neck coil and a GE Healthcare SIGNA Pioneer system equipped with a 32-channel high-resolution head coil. Conventional MRI included axial T1WI, axial T2WI, T2-FLAIR, diffusion-weighted imaging for ADC maps, and high-resolution contrast-enhanced 3D T1WI. DSC-PWI was acquired using gradient-echo echo-planar imaging. Detailed acquisition parameters are summarized in Supplementary Table 1. Scanner vendor and protocol variability were considered when interpreting the generalizability of radiomic, ADC, and rCBV measurements.

Images were converted from Digital Imaging and Communications in Medicine format to Neuroimaging Informatics Technology Initiative format. Skull stripping was performed with the Swiss Skull Stripper plugin in 3D Slicer (version available at the time of analysis; RRID: SCR_005619), and images were registered to the Montreal Neurological Institute 152 template using SlicerANTs. Bias-field correction used the N4 algorithm in SimpleITK with an Otsu-generated mask, 200 histogram bins, maximum iterations of 50, 50, 30, and 20 across four resolution levels, and a convergence threshold of 1e-6. Intensity normalization was performed using the WhiteStripe package in R on skull-stripped images after N4 correction, with the image-type parameter set according to MRI weighting. Images were then resampled to isotropic 1 x 1 x 1 mm voxels using SimpleITK B-spline interpolation before radiomic feature extraction (24–30).

Region-of-interest delineation and radiomic feature extraction

Two radiologists independently delineated the enhancing tumor region of interest on contrast-enhanced T1-weighted imaging (T1CE). The T1CE-derived ROI was then registered to the corresponding T1-weighted, T2-weighted, and T2-fluid-attenuated inversion recovery sequences for multisequence radiomic feature extraction. Interobserver reproducibility was assessed by calculating intraclass correlation coefficients from features extracted from the two independent delineations. Discrepancies were resolved by consensus, and the final consensus ROI was used for model development. During delineation and consensus review, readers were not allowed to use radiomic features or model outputs for outcome assignment.

Radiomic features were extracted using PyRadiomics (version 3.0; Figure 1) from the T1CE-defined enhancing-tumor ROI registered to each sequence. The extraction settings were derived from the original feature-extraction script: fixed bin width was set to 25, the ROI label value was 1, resampledPixelSpacing was set to None, and interpolateResampling was set to False because images had already been resampled before feature extraction. Original and Wavelet image types were enabled; diagnostic metadata returned by PyRadiomics was removed before analysis. Shape, first-order intensity, and higher-order texture features, including gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, gray-level dependence matrix, and neighboring gray-tone difference matrix features, were extracted.

Figure 1.

Flowchart diagram illustrating a radiomics workflow for brain MRI, showing steps including skull stripping, brain registration, bias correction, expert region of interest delineation on 3D brain models, feature extraction, and integration of imaging and clinical data for analysis.

Open in a new tab

Radiomics workflow for extracting magnetic resonance imaging features in glioma. The workflow included image preprocessing, region-of-interest delineation, radiomic feature extraction, reproducibility assessment, and integration with clinical, molecular, diffusion, and perfusion variables. ADC, apparent diffusion coefficient; rCBV, relative cerebral blood volume; ROI, region of interest.

Feature selection and RadScore construction

Radiomic features from the four MRI sequences were analyzed using a strict split-first workflow to prevent overfitting and information leakage. The dataset was first partitioned into training and validation cohorts in a stratified 7:3 ratio. All subsequent radiomics steps were restricted to the training cohort: interobserver reproducibility was assessed using ICCs calculated from the two independent reader matrices, stable features were retained using an ICC threshold of 0.80 or greater (Figure 2), Z-score normalization parameters were derived only from the training cohort, and LASSO logistic regression with 10-fold cross-validation was fitted only in the training cohort. The locked stable-feature list, training-derived means and standard deviations, LASSO-selected features, and LASSO coefficients were then applied unchanged to the validation cohort.

Figure 2.

Multi-panel histogram figure visualizing radiomic feature stability across four MRI modalities: T1, T1CE, T2, and T2FLAIR. Each histogram displays frequency density against the intraclass correlation coefficient (ICC), with a vertical dashed line at ICC 0.8 indicating the reliability threshold. Blue bars represent feature distributions, while red lines show density estimates. Most features in T1CE and T2FLAIR are above the threshold, while T1 and T2 show greater spread below and above the threshold. Total features analyzed is three thousand four hundred four.

Open in a new tab

Intraclass correlation coefficient analysis for radiomic feature stability across magnetic resonance imaging modalities. Features with intraclass correlation coefficients greater than 0.80 were retained for downstream modeling.

ADC and rCBV measurement

ADC was calculated from diffusion-weighted imaging acquired with b values of 0 and 1,000 s/mm2. ADC maps were generated with the exponential decay model S(b) = S (0) x exp(-b x ADC), and mean ADC was measured within the tumor-enhancement region of interest at the first suspected radiological progression after radiotherapy and expressed as x10^-3 mm2/s (31). During ADC measurement, regions of necrosis, hemorrhage, and large vessels were avoided whenever identifiable.

rCBV was derived from DSC-PWI obtained at the first suspected radiological progression after radiotherapy. DSC-PWI consisted of 60 dynamic phases with a temporal resolution of approximately 1.5 seconds. A preload dose of 0.05 mmol/kg gadolinium-based contrast agent was administered approximately 5 minutes before perfusion imaging to reduce T1-leakage effects. During dynamic acquisition, a main bolus dose of 0.1 mmol/kg gadolinium-based contrast agent was injected at approximately the 10th phase using a power injector at 4.5 mL/s, followed by a 20-mL saline flush. Post-processing was performed on Syngo.Via or GE AW workstations using singular value decomposition, and bidirectional T1 and T2* leakage correction based on the Boxerman-Schmainda model was applied to calculate corrected rCBV. rCBV was defined as cerebral blood volume in the tumor region of interest divided by cerebral blood volume in contralateral normal-appearing white matter (32). Regions of necrosis, hemorrhage, and large vessels were avoided during ROI placement.

Model development and statistical analysis

Continuous variables were summarized as median [interquartile range], and categorical variables were summarized as number (percentage). Between-group comparisons used the independent t test or Mann-Whitney U test for continuous variables, as appropriate, and the chi-square test or Fisher exact test for categorical variables. Two-sided P values less than 0.05 were considered statistically significant.

Continuous predictors were standardized before regression modeling unless a clinically interpretable scale was retained, and categorical predictors were entered as indicator variables. Univariable logistic regression was used to screen candidate predictors. Variables with P < 0.05 in univariable analysis were then entered into multivariable logistic regression to identify independent predictors and construct the final model. A nomogram was derived from the multivariable model using the rms package in R.

Model performance was evaluated across three domains: discrimination, calibration, and clinical utility. Discrimination was quantified using receiver operating characteristic curves and area under the curve with 95% confidence intervals. Calibration was assessed with calibration plots, calibration-in-the-large, calibration slope, concordance index, and Brier score. Clinical utility was examined using decision curve analysis. For the leakage-controlled nested analysis, a stratified split was performed with a fixed seed (2026): 70% of patients were assigned to a training cohort and 30% to a held-out validation cohort, preserving the proportion of PsP events. ICC filtering, Z-score normalization, LASSO feature selection, RadScore construction, and integrated-model coefficient estimation were all performed exclusively in the training cohort. The locked workflow was then applied to the validation cohort without feature reselection, parameter updating, or recalibration. Validation-set performance was compared across the integrated model, a RadScore-only model, and a clinical-imaging model that excluded RadScore and retained rCBV, ADC, NLR, MGMT promoter methylation status, and TMZ treatment.

Study size was determined by the number of eligible patients in the retrospective cohort. The full analytic cohort included 222 patients and 60 PsP events. In the split-sample analysis, the training cohort included 155 patients with 42 PsP events, and the validation cohort included 67 patients with 18 PsP events. Because the final multivariable model included six predictors, the split-sample analysis was treated as supportive internal validation, given the reduced statistical efficiency of data splitting in modest-sized datasets.

Results

Patient flow and baseline characteristics

Among 254 screened patients, 32 were excluded because of incomplete required data, inadequate follow-up information, insufficient image quality, or indeterminate post-treatment clinical-radiological evolution. The final complete-case analytic cohort included 222 patients, of whom 60 (27.0%) had PsP and 162 (73.0%) were classified as non-PsP. Among excluded patients, 21 had unresolved post-treatment evolution and were excluded before model development. After complete-case exclusion, no missing values remained for candidate predictors or the outcome. The patient selection process is shown in Figure 3.

Figure 3.

Flowchart illustrating patient selection for a high-grade glioma study: 254 assessed, 32 excluded, 222 included, divided into pseudoprogression (60) and non-pseudoprogression (162) after surgery and radiotherapy with MRI.

Open in a new tab

Flowchart of patient selection. A total of 254 patients were assessed for eligibility. Thirty-two patients were excluded before final analysis, including 21 patients with indeterminate clinical-radiological evolution or unresolved PsP/non-PsP status and 11 patients with incomplete required data, inadequate follow-up, or poor image quality. The final complete-case analytic cohort included 222 patients, comprising 60 patients with pseudoprogression and 162 patients without pseudoprogression.

Baseline characteristics according to PsP status are summarized in Table 1. The groups differed significantly in seizure at presentation, isocitrate dehydrogenase 1 status, MGMT promoter methylation status, TMZ treatment, NLR, rCBV, and ADC.

Table 1.

Baseline characteristics of patients stratified by pseudoprogression status.

Variable Category Non-PsP (n=162) PsP (n=60) P value
Sex Female 70 (43.2%) 23 (38.3%) 0.616
Male 92 (56.8%) 37 (61.7%)
Age (years) 48.0 [35.0, 59.8] 49.0 [38.0, 56.8] 0.678
ECOG performance status 0 52 (32.1%) 13 (21.7%) 0.314
1 69 (42.6%) 30 (50.0%)
2 41 (25.3%) 17 (28.3%)
Extent of resection Biopsy 6 (3.7%) 2 (3.3%) 0.426
Gross total resection 12 (7.4%) 8 (13.3%)
Partial resection 44 (27.2%) 19 (31.7%)
Subtotal resection 100 (61.7%) 31 (51.7%)
RT dose (Gy) 60.2 [54.0, 64.3] 59.8 [54.0, 64.1] 0.985
RT technique Gamma Knife 26 (16.0%) 12 (20.0%) 0.526
IMRT 35 (21.6%) 14 (23.3%)
TOMO 38 (23.5%) 17 (28.3%)
VMAT 63 (38.9%) 17 (28.3%)
Tumor volume (cm3) 86.0 [47.3, 165.2] 91.2 [45.9, 174.6] 0.888
Edema Indeterminate 33 (20.4%) 11 (18.3%) 0.866
No 91 (56.2%) 33 (55.0%)
Yes 38 (23.5%) 16 (26.7%)
Location Cortical 101 (62.3%) 39 (65.0%) 0.807
Deep 29 (17.9%) 8 (13.3%)
Midline/ventricular 12 (7.4%) 6 (10.0%)
Posterior/brainstem 20 (12.3%) 7 (11.7%)
Seizure Yes 34 (21.0%) 21 (35.0%) 0.049
No 128 (79.0%) 39 (65.0%)
IDH1 Mutant 100 (61.7%) 27 (45.0%) 0.037
Wild-type 62 (38.3%) 33 (55.0%)
MGMT promoter Unmethylated/negative 51 (31.5%) 6 (10.0%) 0.002
Methylated/positive 111 (68.5%) 54 (90.0%)
TMZ Yes 53 (32.7%) 34 (56.7%) 0.002
No 109 (67.3%) 26 (43.3%)
NLR 2.3 [1.4, 2.9] 3.3 [2.2, 3.8] <0.001
rCBV 2.1 [1.2, 2.8] 2.8 [2.0, 3.3] <0.001
ADC 1.4 [1.1, 1.6] 1.0 [0.9, 1.3] <0.001
Dexamethasone dose (mg) 32.5 [20.0, 43.8] 35.0 [25.0, 50.0] 0.241
Open in a new tab

Data are presented as median [interquartile range] for continuous variables and number (percentage) for categorical variables. ADC, apparent diffusion coefficient; ECOG, Eastern Cooperative Oncology Group; IDH1, isocitrate dehydrogenase 1; IMRT, intensity-modulated radiotherapy; MGMT, O6-methylguanine-DNA methyltransferase; NLR, neutrophil-to-lymphocyte ratio; PsP, pseudoprogression; rCBV, relative cerebral blood volume; RT, radiotherapy; TMZ, temozolomide; TOMO, tomotherapy; VMAT, volumetric modulated arc therapy.

Radiomic feature selection and RadScore performance

For each patient, 3,404 radiomic features were analyzed from the four MRI sequences, representing 851 features from each of the T1WI, T1CE, T2WI, and T2-FLAIR sequences. Training-only ICC filtering retained 132 stable features. After training-set Z-score normalization, LASSO logistic regression with 10-fold cross-validation selected 17 non-zero-coefficient features (Figure 4): T1CE_original_firstorder_Entropy, T1CE_original_ngtdm_Complexity, T2_original_firstorder_Mean, T2_original_glcm_Idmn, T2FLAIR_original_glszm_ZoneEntropy, T1CE_original_firstorder_Minimum, T1CE_original_shape_MeshVolume, T2FLAIR_original_glszm_SizeZoneNonUniformity, T2FLAIR_original_gldm_SmallDependenceEmphasis, T1_original_ngtdm_Busyness, T2_original_glszm_ZonePercentage, T2FLAIR_original_glrlm_HighGrayLevelRunEmphasis, T1CE_original_shape_Maximum3DDiameter, T2FLAIR_original_glrlm_GrayLevelVariance, T2_original_glcm_MaximumProbability, T2_original_glcm_Idn, and T2FLAIR_original_glrlm_LongRunLowGrayLevelEmphasis. Their coefficients are shown in Figure 5.

Figure 4.

Panel A shows a line plot of LASSO coefficient paths for multiple variables, with coefficients on the y-axis and log(lambda) on the x-axis, and a vertical dashed line indicating lambda selection. Panel B presents a cross-validation plot with binomial deviance on the y-axis and log(lambda) on the x-axis, displaying mean deviance values with error bars and a vertical dashed line for lambda selection.

Open in a new tab

Radiomic feature selection using least absolute shrinkage and selection operator regression. (A) Coefficient path plot. (B) Cross-validation error plot used to select the optimal regularization parameter.

Figure 5.

Bar chart and pie chart presenting radiomic feature selection by LASSO and MRI sequence contribution. Bar chart shows coefficients for various features, both positive and negative, color-coded by MRI sequence. Pie chart indicates T2FLAIR contributes the most features at 35.3 percent, followed by T1CE and T2 at 29.4 percent each, and T1 at 5.9 percent.

Open in a new tab

Contribution of selected radiomic features. (A) Bar plot of absolute least absolute shrinkage and selection operator coefficients for the 17 retained features. (B) Pie chart showing the relative contribution of magnetic resonance imaging sequences to the radiomics score.

In the leakage-controlled split-first analysis, the RadScore-alone model showed moderate and relatively stable discrimination. Its AUC was 0.890 (95% CI: 0.838-0.942) in the training cohort and 0.771 (95% CI: 0.648-0.894) in the validation cohort, suggesting that the nested radiomics signature retained predictive information after leakage-controlled internal validation. The distribution of the locked nested RadScore according to PsP status in the training and validation cohorts is shown in Figure 6.

Figure 6.

Box plot illustrating Nested RadScore distributions for Non-PsP and PsP groups across Training and Validation datasets, with individual data points overlaid; PsP groups show higher median scores in both datasets.

Open in a new tab

Distribution of the locked nested RadScore in the training and validation cohorts stratified by pseudoprogression status. Boxplots with overlaid individual points show RadScore values for non-PsP and PsP patients in each cohort. PsP, pseudoprogression.

Clinical predictors and final model specification

In univariable logistic regression, RadScore, seizure at presentation, isocitrate dehydrogenase 1 status, MGMT promoter methylation, TMZ treatment, NLR, rCBV, ADC, and dexamethasone dose were associated with PsP (all P < 0.05; Figure 7).

Figure 7.

Forest plot graphic showing odds ratios with 95 percent confidence intervals and p-values for multiple variables, including RadScore, Seizure, IDH1 mutation, MGMT, TMZ, NLR, rCBV, ADC, and Dexamethasone, with significant results identified for several factors.

Open in a new tab

Forest plot of univariable logistic regression analysis for candidate predictors of pseudoprogression. CI, confidence interval; OR, odds ratio.

Multivariable logistic regression retained RadScore, rCBV, ADC, NLR, MGMT promoter methylation, and TMZ treatment as independent predictors of PsP. The adjusted results were RadScore (odds ratio, 1.31; 95% confidence interval: 1.21-1.45; P < 0.001), rCBV (odds ratio, 3.06; 95% confidence interval: 1.66-6.18; P < 0.001), ADC (odds ratio, 0.07; 95% confidence interval: 0.01-0.52; P = 0.012), NLR (odds ratio, 1.80; 95% confidence interval: 1.04-3.18; P = 0.036), MGMT promoter methylation positive status (odds ratio, 4.48; 95% confidence interval: 1.16-21.28; P = 0.040), and TMZ treatment (odds ratio, 3.10; 95% confidence interval: 1.03-10.21; P = 0.049; Figure 8).

Figure 8.

Forest plot showing odds ratios with ninety-five percent confidence intervals for six variables: RadScore, rCBV, ADC, NLR, MGMT plus, and TMZ Yes. Corresponding odds ratios, confidence intervals, and p values are summarized in columns.

Open in a new tab

Forest plot of multivariable logistic regression analysis for independent predictors of pseudoprogression. CI, confidence interval; OR, odds ratio.

The final logistic regression equation was Logit(PsP) = -3.971 + 0.269 x RadScore + 1.174 x rCBV - 2.838 x ADC + 0.617 x NLR + 1.305 x MGMT(+) + 1.034 x TMZ(Yes). The predicted probability was calculated as P(PsP) = 1/[1 + exp(-Logit(PsP))]. MGMT(+) was coded as 1 for positive MGMT promoter methylation and 0 for negative methylation; TMZ(Yes) was coded as 1 for receipt of TMZ treatment and 0 for no TMZ treatment. The corresponding nomogram is shown in Figure 9.

Figure 9.

Nomogram for predicting PsP in glioma displaying multiple horizontal line plots corresponding to ADC, rCBV, NLR, TMZ, MGMT, and RadScore variables, with point scales above and a total points scale below, indicating probability values for clinical interpretation.

Open in a new tab

Nomogram for individualized prediction of pseudoprogression after radiotherapy in patients with high-grade glioma.

Model performance

In the leakage-controlled split-first analysis, the integrated model showed an AUC of 0.898 (95% CI: 0.850-0.946) and a Brier score of 0.116 in the training cohort. In the held-out validation cohort, the integrated model achieved an AUC of 0.811 (95% CI: 0.696-0.925), a Brier score of 0.159, a mean predicted versus observed PsP rate of 0.268 versus 0.269, a calibration intercept of 0.006, and a calibration slope of 0.575. The clinical-imaging model without RadScore achieved a validation AUC of 0.744 (95% CI: 0.614-0.873) and a Brier score of 0.178, whereas RadScore alone achieved a validation AUC of 0.771 (95% CI: 0.648-0.894) and a Brier score of 0.185. Detailed performance metrics for the integrated, clinical-imaging, and RadScore-only models in the training and held-out validation cohorts are summarized in Table 2. The validation ROC curves, calibration plots, and decision curve analysis for the integrated model, clinical-imaging model, and RadScore-only model are shown in Figure 10.

Table 2.

Performance of the integrated, clinical-imaging, and RadScore-only models in the stratified 7:3 split-sample internal validation analysis.

Model Cohort n/events AUC (95% CI) Brier score Mean predicted/observed Calib. intercept Calib. slope
Integrated Training 155/42 0.898 (0.850-0.946) 0.116 0.271 vs 0.271 0.000 1.000
Integrated Validation 67/18 0.811 (0.696-0.925) 0.159 0.268 vs 0.269 0.006 0.575
Clinical-imaging Training 155/42 0.863 (0.804-0.923) 0.133 0.271 vs 0.271 -0.000 1.000
Clinical-imaging Validation 67/18 0.744 (0.614-0.873) 0.178 0.263 vs 0.269 0.043 0.589
RadScore only Training 155/42 0.890 (0.838-0.942) 0.118 0.271 vs 0.271 0.000 1.000
RadScore only Validation 67/18 0.771 (0.648-0.894) 0.185 0.268 vs 0.269 0.011 0.411
Open in a new tab

AUC, area under the receiver operating characteristic curve; CI, confidence interval. The leakage-controlled split-first workflow used 3,404 candidate radiomic features from complete two-reader matrices. Dataset partitioning preceded ICC filtering, Z-score normalization, and LASSO feature selection. The locked RadScore was constructed from LASSO-selected radiomic features. The integrated model included RadScore, rCBV, ADC, NLR, MGMT promoter methylation status, and temozolomide treatment; the clinical-imaging model excluded RadScore. Each model was fitted in the training cohort and applied to the validation cohort without recalibration.

Figure 10.

Panel A displays a ROC curve comparing Clinical (orange, AUC 0.74), RadScore (blue, AUC 0.77), and Integrated (green, AUC 0.81) models for sensitivity versus one minus specificity. Panel B presents a calibration plot showing observed versus predicted PsP rate for the same three models, each represented by circles joined by lines. Panel C features a decision curve analysis plotting net benefit against threshold probability, with colored lines for Integrated, Clinical, RadScore, and gray dashed lines for “Treat all” and “Treat none” strategies.

Open in a new tab

Split-sample internal validation and model comparison. (A) Receiver operating characteristic curves in the held-out validation cohort comparing the integrated model, RadScore-only model, and clinical-imaging model. (B) Calibration plot showing agreement between predicted and observed pseudoprogression probabilities. (C) Decision curve analysis comparing the integrated model, RadScore-only model, clinical-imaging model, treat-all strategy, and treat-none strategy. AUC, area under the receiver operating characteristic curve; PsP, pseudoprogression.

Discussion

This study developed and internally validated a multimodal prediction model for PsP after radiotherapy in patients with HGG. The final model integrated radiomic, perfusion, diffusion, inflammatory, molecular, and treatment-related information. RadScore, rCBV, ADC, NLR, MGMT promoter methylation, and TMZ treatment were independently associated with PsP. The integrated model showed high apparent discrimination, acceptable calibration, and favorable decision-curve net benefit in the development cohort, with maintained discrimination in the held-out internal validation cohort.

Most of the 17 radiomic features selected by least absolute shrinkage and selection operator regression were texture features, particularly from contrast-enhanced T1-weighted and T2-fluid-attenuated inversion recovery sequences. T1CE_original_firstorder_Entropy was the strongest positive contributor to RadScore, suggesting that signal disorder within enhancing tissue may reflect post-radiotherapy inflammation, vascular permeability, necrosis, and tissue disruption (33–37). Features capturing spatial complexity, zone entropy, and neighborhood gray-tone variation may further quantify subtle tissue patterns that are difficult to appreciate visually.

The independent associations of perfusion and diffusion parameters with PsP underscore the biological complexity of post-treatment enhancement. In many studies, TP is associated with neovascularization and higher perfusion, whereas PsP tends to show lower perfusion. The positive association between rCBV and PsP in this cohort should therefore be interpreted cautiously. One possible explanation is that early treatment-related injury may produce inflammatory hyperemia, severe blood-brain barrier disruption, and reactive permeability changes that can influence apparent rCBV estimates even when preload dosing and leakage correction are applied. Differences in imaging timing, ROI placement, scanner platform, and normalization strategy may also contribute to this observation (6, 38). The association between lower ADC and PsP should likewise be interpreted as hypothesis-generating; lower ADC may reflect transient hypercellularity related to inflammatory infiltration, cytotoxic edema, vascular injury, necrosis, or temporal variation in diffusion characteristics after treatment rather than viable tumor alone.

NLR was independently associated with PsP, supporting a potential contribution of systemic inflammation to post-radiotherapy imaging changes. Elevated NLR may reflect a pro-inflammatory state and altered immune balance, both of which could plausibly contribute to radiation-related inflammatory responses (39). MGMT promoter methylation and TMZ treatment were also independent predictors. MGMT promoter methylation is associated with sensitivity to alkylating therapy and has repeatedly been linked to higher rates of PsP, whereas TMZ combined with radiotherapy may intensify treatment-related injury and inflammatory response (40, 41). Recent methodological work on brain tumor segmentation, radiomics reproducibility, ICC reliability, penalized regression, clinical predictors, and calibrated radiomics models further supports the need for standardized ROI definition, robustness assessment, and calibration/decision-curve evaluation in glioma imaging studies (42–51).

Incorporating the RadScore into the clinical-imaging baseline improved the validation AUC from 0.744 to 0.811 and reduced the Brier score from 0.178 to 0.159. These findings support the complementary value of radiomic heterogeneity beyond conventional clinical and advanced imaging variables. The integrated model also provides a clinically interpretable multimodal framework by jointly considering radiomic, perfusion-diffusion, inflammatory, molecular, and treatment-related information. Given the modest validation sample size and single-center design, these comparative results should be interpreted cautiously and require external confirmation.

Several limitations should be acknowledged. First, although 13 patients classified as PsP underwent repeat surgery or biopsy with histopathological assessment, the endpoint for most patients was based on RANO/RANO 2.0-guided longitudinal clinical-radiological assessment rather than uniform tissue confirmation. To reduce misclassification, non-histologically confirmed PsP required at least 6 months of stable or improved follow-up imaging, outcome adjudication was performed through multidisciplinary review by two radiologists and two medical oncologists, and 21 indeterminate cases were excluded before model development. Second, the retrospective single-center design may introduce selection bias and limit generalizability. Third, although the leakage-controlled split-first workflow reduced the risk of information leakage, the validation cohort was relatively small and derived from the same institution; therefore, model calibration and comparative performance among the integrated, clinical-imaging, and RadScore-only models should be interpreted cautiously. Fourth, no independent external validation cohort was available; therefore, all performance estimates remain internally validated only. Fifth, complete-case analysis may introduce bias if excluded patients differed systematically from included patients. Sixth, scanner hardware, acquisition protocol, image preprocessing, ADC/rCBV measurement, and ROI delineation can affect quantitative imaging features; ICC filtering reduces segmentation-related instability but does not fully address acquisition-related heterogeneity. Prospective multicenter validation with standardized imaging protocols, locked radiomics code, complete raw feature matrices, and, where feasible, histopathological or rigorously adjudicated reference standards is required.

Conclusion

In conclusion, this single-center exploratory study developed a leakage-controlled split-first prediction workflow for PsP after radiotherapy in patients with HGG. The nested RadScore retained validation-set discriminative information and modestly improved the clinical-imaging model when incorporated into an integrated multimodal framework. This model may support individualized PsP risk estimation by combining radiomic, perfusion-diffusion, inflammatory, molecular, and treatment-related information, but external validation with standardized imaging protocols and locked model coefficients is required before clinical implementation.

Acknowledgments

The authors would like to thank all patients who participated in the study.

Funding Statement

The author(s) declared that financial support was received for this work and/or its publication. This work was supported by the Liuzhou Science and Technology Plan Project (Grant No. 2022SB011).

Footnotes

Edited by: Min Wu, Sichuan University, China

Reviewed by: Eryi Sun, Jiangsu University Affiliated People’s Hospital, China

Rafail Christodoulou, Stanford University, United States

Deng Xiong, Stevens Institute of Technology, United States

Data availability statement

The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request, subject to patient privacy protection, institutional policies, and approval by the relevant ethics committee where required.

Ethics statement

The studies involving humans were approved by The Medical Research Ethics Committee of Liuzhou Workers’ Hospital. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin because the requirement for written informed consent was waived by the ethics committee because of the retrospective study design.

Author contributions

JZ: Conceptualization, Methodology, Formal analysis, Writing – original draft, Writing – review & editing. LY: Conceptualization, Writing – review & editing. YH: Methodology, Writing – review & editing. ZD: Investigation, Data curation, Writing – review & editing. XZ: Formal analysis, Writing – review & editing. WM: Formal analysis, Writing – review & editing.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fonc.2026.1892760/full#supplementary-material

DataSheet1.docx (37.1KB, docx)

References

  • 1. Karremann M, Gielen GH, Hoffmann M, Wiese M, Colditz N, Warmuth-Metz M, et al. Diffuse high-grade gliomas with H3 K27M mutations carry a dismal prognosis independent of tumor location. Neuro Oncol. (2018) 20:123–31. doi:  10.1093/neuonc/nox149 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Omuro A, Brandes AA, Carpentier AF, Idbaih A, Reardon DA, Cloughesy T, et al. Radiotherapy combined with nivolumab or temozolomide for newly diagnosed glioblastoma with unmethylated MGMT promoter: An international randomized phase III trial. Neuro Oncol. (2023) 25:123–34. doi:  10.1093/neuonc/noac099 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Mekala JR, Adusumilli K, Chamarthy S, Angirekula HSR. Novel sights on therapeutic, prognostic, and diagnostics aspects of non-coding RNAs in glioblastoma multiforme. Metab Brain Dis. (2023) 38:1801–29. doi:  10.1007/s11011-023-01234-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Ali M, Fulci G, Grigalavicius M, Pulli B, Li A, Wojtkiewicz GR, et al. Myeloperoxidase exerts anti-tumor activity in glioma after radiotherapy. Neoplasia. (2022) 26:100779. doi:  10.1016/j.neo.2022.100779 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Amer A, Ansari S, Krayyem A, Kundu S, Khose S, Pokhylevych H, et al. Dynamic contrast-enhanced MRI processing comparison for distinguishing true progression from pseudoprogression in high-grade glioma. J Comput Assist Tomogr. (2025) 49(4):656–61. doi:  10.1097/RCT.0000000000001716 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Sidibe I, Tensaouti F, Roques M, Cohen-Jonathan-Moyal E, Laprie A. Pseudoprogression in glioblastoma: Role of metabolic and functional MRI-systematic review. Biomedicines. (2022) 10(2):285. doi:  10.3390/biomedicines10020285 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Knudsen-Baas KM, Moen G, Fluge O, Storstein A. Pseudoprogression in high-grade glioma. Acta Neurol Scand Suppl. (2013) 196:31–7. doi:  10.1111/ane.12047 [DOI] [PubMed] [Google Scholar]
  • 8. Abbasi AW, Westerlaan HE, Holtman GA, Aden KM, van Laar PJ, van der Hoorn A, et al. Incidence of tumour progression and pseudoprogression in high-grade gliomas: A systematic review and meta-analysis. Clin Neuroradiol. (2018) 28:401–11. doi:  10.1007/s00062-017-0584-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Brandes AA, Franceschi E, Tosoni A, Blatt V, Pession A, Tallini G, et al. MGMT promoter methylation status can predict the incidence and outcome of pseudoprogression after concomitant radiochemotherapy in newly diagnosed glioblastoma patients. J Clin Oncol. (2008) 26:2192–7. doi:  10.1200/JCO.2007.14.8163 [DOI] [PubMed] [Google Scholar]
  • 10. Stumpo V, Guida L, Bellomo J, Van Niftrik CHB, Sebok M, Berhouma M, et al. Hemodynamic imaging in cerebral diffuse glioma-part B: Molecular correlates, treatment effect monitoring, prognosis, and future directions. Cancers (Basel). (2022) 14(5):1342. doi:  10.3390/cancers14051342 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Wen PY, van den Bent M, Youssef G, Cloughesy TF, Ellingson BM, Weller M, et al. RANO 2.0: Update to the response assessment in neuro-oncology criteria for high- and low-grade gliomas in adults. J Clin Oncol. (2023) 41:5187–99. doi:  10.1200/JCO.23.01059 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Moon HH, Kim HS, Park JE, Kim YH, Kim JH. Refinement of response assessment in neuro-oncology (RANO) using non-enhancing lesion type and contrast enhancement evolution pattern in IDH wild-type glioblastomas. BMC Cancer. (2021) 21:654. doi:  10.1186/s12885-021-08414-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Overcast WB, Davis KM, Ho CY, Hutchins GD, Green MA, Graner BD, et al. Advanced imaging techniques for neuro-oncologic tumor diagnosis, with an emphasis on PET-MRI imaging of Malignant brain tumors. Curr Oncol Rep. (2021) 23:34. doi:  10.1007/s11912-021-01020-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Gu X, He X, Wang H, Li J, Chen R, Liu H, et al. Dynamic susceptibility contrast-enhanced perfusion-weighted imaging in differentiation between recurrence and pseudoprogression in high-grade glioma: A meta-analysis. J Comput Assist Tomogr. (2024) 48:303–10. doi:  10.1097/RCT.0000000000001543 [DOI] [PubMed] [Google Scholar]
  • 15. Yu Y, Ren W, He Z, Chen Y, Tan Y, Mao L, et al. Machine learning radiomics of magnetic resonance imaging predicts recurrence-free survival after surgery and correlation of LncRNAs in patients with breast cancer: A multicenter cohort study. Breast Cancer Res. (2023) 25:132. doi:  10.1186/s13058-023-01688-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Balana C, Castaner S, Carrato C, Moran T, Lopez-Paradis A, Domenech M, et al. Preoperative diagnosis and molecular characterization of gliomas with liquid biopsy and radiogenomics. Front Neurol. (2022) 13:865171. doi:  10.3389/fneur.2022.865171 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Fang S, Yang Y, Xu N, Tu Y, Yin Z, Zhang Y, et al. An update in imaging evaluation of histopathological grade of soft tissue sarcomas using structural and quantitative imaging and radiomics. J Magn Reson Imaging. (2022) 55:1357–75. doi:  10.1002/jmri.27954 [DOI] [PubMed] [Google Scholar]
  • 18. Conti A, Duggento A, Indovina I, Guerrisi M, Toschi N. Radiomics in breast cancer classification and prediction. Semin Cancer Biol. (2021) 72:238–50. doi:  10.1016/j.semcancer.2020.04.002 [DOI] [PubMed] [Google Scholar]
  • 19. Yan J, Zhang B, Zhang S, Cheng J, Liu X, Wang W, et al. Quantitative MRI-based radiomics for noninvasively predicting molecular subtypes and survival in glioma patients. NPJ Precis Oncol. (2021) 5:72. doi:  10.1038/s41698-021-00205-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Li J, Xu Q, Fan X, Cheng X, Tao J, Lu H, et al. Multi-sequence MRI radiomics model for discrimination of recurrence and pseudoprogression in gliomas. Eur J Radiol. (2025) 194:112508. doi:  10.1016/j.ejrad.2025.112508 [DOI] [PubMed] [Google Scholar]
  • 21. Li K, Zhu Q, Yang J, Zheng Y, Du S, Song M, et al. Imaging and liquid biopsy for distinguishing true progression from pseudoprogression in gliomas, current advances and challenges. Acad Radiol. (2024) 31:3366–83. doi:  10.1016/j.acra.2024.03.019 [DOI] [PubMed] [Google Scholar]
  • 22. Lohmann P, Elahmadawy MA, Gutsche R, Werner JM, Bauer EK, Ceccon G, et al. FET PET radiomics for differentiating pseudoprogression from early tumor progression in glioma patients post-chemoradiation. Cancers (Basel). (2020) 12(12):3835. doi:  10.3390/cancers12123835 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Reddy S, Lung T, Muniyappa S, Hadley C, Templeton B, Fritz J, et al. Radiomics and radiogenomics in differentiating progression, pseudoprogression, and radiation necrosis in gliomas. Biomedicines. (2025) 13(7):1778. doi:  10.3390/biomedicines13071778 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Li X, Morgan PS, Ashburner J, Smith J, Rorden C. The first step for neuroimaging data analysis: DICOM to NIfTI conversion. J Neurosci Methods. (2016) 264:47–56. doi:  10.1016/j.jneumeth.2016.03.001 [DOI] [PubMed] [Google Scholar]
  • 25. Pei L, Ak M, Tahon NHM, Zenkin S, Alkarawi S, Kamal A, et al. A general skull stripping of multiparametric brain MRIs using 3D convolutional neural network. Sci Rep. (2022) 12:10826. doi:  10.1038/s41598-022-14983-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Giff A, Noren G, Magnotti J, Lopes AC, Batistuzzo MC, Hoexter M, et al. Spatial normalization discrepancies between native and MNI152 brain template scans in gamma ventral capsulotomy patients. Psychiatry Res Neuroimaging. (2023) 329:111595. doi:  10.1016/j.pscychresns.2023.111595 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Kanakaraj P, Yao T, Cai LY, Lee HH, Newlin NR, Kim ME, et al. DeepN4: Learning N4ITK bias field correction for T1-weighted images. Neuroinformatics. (2024) 22:193–205. doi:  10.1007/s12021-024-09655-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, et al. N4ITK: Improved N3 bias correction. IEEE Trans Med Imaging. (2010) 29:1310–20. doi:  10.1109/TMI.2010.2046908 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Ghassemi R, Brown R, Narayanan S, Banwell B, Nakamura K, Arnold DL, et al. Normalization of white matter intensity on T1-weighted images of patients with acquired central nervous system demyelination. J Neuroimaging. (2015) 25:184–90. doi:  10.1111/jon.12129 [DOI] [PubMed] [Google Scholar]
  • 30. McMaster EM, Newlin NR, Rudravaram G, Saunders AM, Krishnan AR, Remedios LW, et al. Harmonized connectome resampling for variance in voxel sizes. Magn Reson Imaging. (2025) 121:110424. doi:  10.1016/j.mri.2025.110424 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Sahoo P, Rockne RC, Jung A, Gupta PK, Rathore RKS, Gupta RK, et al. Synthetic apparent diffusion coefficient for high b-value diffusion-weighted MRI in prostate. Prostate Cancer. (2020) 2020:5091218. doi:  10.1155/2020/5091218 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Korfiatis P, Kline TL, Kelm ZS, Carter RE, Hu LS, Erickson BJ, et al. Dynamic susceptibility contrast-MRI quantification software tool: Development and evaluation. Tomography. (2016) 2:448–56. doi:  10.18383/j.tom.2016.00172 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Le Fevre C, Constans JM, Chambrelant I, Antoni D, Bund C, Leroy-Freschini B, et al. Pseudoprogression versus true progression in glioblastoma patients: A multiapproach literature review. Part 2 - radiological features and metric markers. Crit Rev Oncol Hematol. (2021) 159:103230. doi:  10.1016/j.critrevonc.2021.103230 [DOI] [PubMed] [Google Scholar]
  • 34. D'Amore F, Grinberg F, Mauler J, Galldiks N, Blazhenets G, Farrher E, et al. Combined 18F-FET PET and diffusion kurtosis MRI in posttreatment glioblastoma: Differentiation of true progression from treatment-related changes. Neuro-Oncol Adv. (2021) 3:vdab044. doi:  10.1093/noajnl/vdab044 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Wang Z, Wang L, Wang Y. Radiomics in glioma: Emerging trends and challenges. Ann Clin Transl Neurol. (2025) 12(3):460–77. doi:  10.1002/acn3.52306 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36. Fu FX, Cai QL, Li G, Wu XJ, Hong L, Chen WS, et al. The efficacy of using a multiparametric magnetic resonance imaging-based radiomics model to distinguish glioma recurrence from pseudoprogression. Magn Reson Imaging. (2024) 111:168–78. doi:  10.1016/j.mri.2024.05.003 [DOI] [PubMed] [Google Scholar]
  • 37. Scalco E, Rizzo G. Texture analysis of medical images for radiotherapy applications. Br J Radiol. (2017) 90:20160642. doi:  10.1259/bjr.20160642 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Ellingson BM, Chung C, Pope WB, Boxerman JL, Kaufmann TJ. Pseudoprogression, radionecrosis, inflammation or true tumor progression? Challenges associated with glioblastoma response assessment in an evolving therapeutic landscape. J Neuro-Oncol. (2017) 134:495–504. doi:  10.1007/s11060-017-2375-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Huang Y, Ding H, Wu Q, Li Z, Li H, Li S, et al. Neutrophil-lymphocyte ratio dynamics are useful for distinguishing between recurrence and pseudoprogression in high-grade gliomas. Cancer Manag Res. (2019) 11:6003–9. doi:  10.2147/CMAR.S202546 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Zhou M, Niu C, Jia L, He H. The value of MGMT promoter methylation and IDH-1 mutation on diagnosis of pseudoprogression in patients with high-grade glioma: A meta-analysis. Med (Baltimore). (2019) 98:e18194. doi:  10.1097/MD.0000000000018194 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41. Li H, Li J, Cheng G, Zhang J, Li X. IDH mutation and MGMT promoter methylation are associated with the pseudoprogression and improved prognosis of glioblastoma multiforme patients who have undergone concurrent and adjuvant temozolomide-based chemoradiotherapy. Clin Neurol Neurosurg. (2016) 151:31–6. doi:  10.1016/j.clineuro.2016.10.004 [DOI] [PubMed] [Google Scholar]
  • 42. Li S, Liu J, Song Z. Brain tumor segmentation based on region of interest-aided localization and segmentation U-Net. Int J Mach Learn Cybern. (2022) 13:2435–45. doi:  10.1007/s13042-022-01536-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Tam LT, Yeom KW, Wright JN, Jaju A, Radmanesh A, Han M, et al. MRI-based radiomics for prognosis of pediatric diffuse intrinsic pontine glioma: An international study. Neuro-Oncol Adv. (2021) 3:vdab042. doi:  10.1093/noajnl/vdab042 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. Schurink NW, van Kranen SR, Roberti S, van Griethuysen JJM, Bogveradze N, Castagnoli F, et al. Sources of variation in multicenter rectal MRI data and their effect on radiomics feature reproducibility. Eur Radiol. (2022) 32:1506–16. doi:  10.1007/s00330-021-08251-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45. Mehta S, Bastero-Caballero RF, Sun Y, Zhu R, Murphy DK, Hardas B, et al. Performance of intraclass correlation coefficient (ICC) as a reliability index under various distributions in scale reliability studies. Stat Med. (2018) 37:2734–52. doi:  10.1002/sim.7679 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Tibshirani R. Regression shrinkage and selection via the Lasso. J R Stat Soc Ser B Methodol. (2018) 58:267–88. doi:  10.1111/j.2517-6161.1996.tb02080.x 40046247 [DOI] [Google Scholar]
  • 47. Le Fevre C, Lhermitte B, Ahle G, Chambrelant I, Cebula H, Antoni D, et al. Pseudoprogression versus true progression in glioblastoma patients: A multiapproach literature review: Part 1 - molecular, morphological and clinical features. Crit Rev Oncol Hematol. (2021) 157:103188. doi:  10.1016/j.critrevonc.2020.103188 [DOI] [PubMed] [Google Scholar]
  • 48. Hoosemans LJN, Panayotopoulos LF, van Kuijk SMJ, Verduin M, Anten MHME, Pasmans RCOS, et al. Clinical predictors of pseudoprogression in glioblastoma: A retrospective cohort analysis. J Neuro-Oncol. (2025) 176:42. doi:  10.1007/s11060-025-05299-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49. Chen L. Overview of clinical prediction models. Ann Transl Med. (2020) 8:71. doi:  10.21037/atm.2019.11.121 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50. Christodoulou RC, Vamvouras G, Pitsillos R, Solomou EE, Georgiou MF. Explainable radiomics with probability calibration for postoperative glioblastoma surveillance. Eur J Radiol Artif Intell. (2026) 5:100074. doi:  10.1016/j.ejrai.2026.100074 38826717 [DOI] [Google Scholar]
  • 51. Christodoulou RC, Vamvouras G, Pitsillos R, Solomou EE, Georgiou MF. Explainable, modality-adaptive radiomics for MGMT methylation prediction in high-grade glioma: a decision-curve analysis study. Front Oncol. (2026) 15:1731258. doi:  10.3389/fonc.2025.1731258 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

DataSheet1.docx (37.1KB, docx)

Data Availability Statement

The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request, subject to patient privacy protection, institutional policies, and approval by the relevant ethics committee where required.

Articles from Frontiers in Oncology are provided here courtesy of Frontiers Media SA

RESOURCES