Integrating ConvNeXt Tiny and Radiomics in a Nomogram to Differentiate True Progression from Pseudoprogression in Glioma

Integrating ConvNeXt Tiny and Radiomics in a Nomogram to Differentiate True Progression from Pseudoprogression in Glioma | 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 Integrating ConvNeXt Tiny and Radiomics in a Nomogram to Differentiate True Progression from Pseudoprogression in Glioma Tao Zheng, Linsha Yang, Duo Zhang, Juan Du, Xin Liang, Shuo Wu, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8643558/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 Differentiating true progression (TP) from pseudoprogression (PsP) in glioma is challenging due to overlapping enhancement patterns on conventional MRI. Therefore, a reliable noninvasive approach integrating imaging heterogeneity is needed to improve TP/PsP discrimination. Methods This multicenter retrospective study included 294 patients with true progression (TP, n = 208) or pseudoprogression (PsP, n = 86). Baseline multiparametric MRI was analyzed. Traditional radiomics and deep learning features extracted using a pre-trained ConvNeXt Tiny network were selected through reproducibility, redundancy, and LASSO analyses to construct imaging signatures, which were combined with clinical factors to develop a deep learning radiomics nomogram (DLRN). Model performance was evaluated using ROC analysis, calibration curves, and decision curve analysis, and compared with radiologists’ assessments. Results The DLRN demonstrated excellent predictive efficacy, achieving an area under the curve (AUC) of 0.873 in the test set. Its performance significantly surpassed that of any individual signature (DeLong test, P < 0.001) and the independent assessments of two senior radiologists. The model exhibited good calibration, and decision curve analysis confirmed its superior clinical net benefit across a wide range of threshold probabilities. When used as a decision-support tool, the nomogram significantly and consistently improved both radiologists' diagnostic performance, yielding a net reclassification improvement greater than 1.1 in both the training and test sets (all P < 0.01). Conclusion The deep learning imaging biomarker nomogram demonstrated excellent performance in differentiating TP from PsP in gliomas, outperforming traditional methods and radiologists, and effectively assisting clinical decision-making. glioma deep learning pseudo-progression true progression Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1 Introduction Glioma is the most common primary malignant tumor in the adult central nervous system, accounting for approximately 80% of all malignant brain tumors. For newly diagnosed high-grade gliomas, the standard treatment includes surgical resection in the maximum safe range, followed by concurrent chemoradiotherapy and adjuvant chemotherapy based on temozolomide( 1 ). Approximately 20%-30% of patients will have new or expanded enhanced lesions that meet the imaging "progression" criteria in the initial follow-up MRI after treatment( 2 ). Some of these lesions represent true tumor progression (TP) and require active treatments, such as reoperation and changes in chemotherapy regimen( 2 ). However, another part is pseudo-progression (PsP), an excessive local tissue response confined to the tumor area subsequent to effective treatment( 3 , 4 ). Currently, the clinical differentiation between TP and PsP in glioma primarily relies on imaging follow-up. However, in the early stage after treatment, both TP and PsP can present similar abnormal enhancement foci and peripheral edema on conventional MRI ( 5 ). Arterial spin-labeled perfusion imaging revealed that the increase in cerebral blood flow was more significant in patients with signs of PsP( 6 ). Positron emission computed tomography (PET-CT) scans also showed that the uptake of specific tracers in patients with recurrent high-grade glioma is significantly higher than that in PsP lesions( 7 ). Despite offering supplementary insights into tumor angiogenesis, cellularity, and metabolism, these advanced functional imaging techniques are limited by non-standardized protocols, high complexity and cost, and restricted routine clinical applicability. Several studies have confirmed that radiomics performs well in differentiating TP from PsP( 8 , 9 ). However, traditional radiomics methods heavily rely on manual delineation of regions of interest, which is not only time-consuming but also introduces inter-observer variability( 10 ). Deep learning technology can automatically extract complex and deep features from raw images through an end-to-end learning approach, overcoming the limitations of manual design and demonstrating stronger feature expression and generalization capabilities. Currently, various advanced architectures, including convolutional neural networks and vision transformers, have achieved significant progress in automatic segmentation of gliomas, non-invasive prediction of isocitrate dehydrogenase (IDH) mutation status, tumor grading, and prognosis assessment( 11 – 13 ). This study aimed to develop and validate a multimodal nomogram to noninvasively differentiate TP from PsP in glioma. Innovatively, we combined deep features extracted automatically by the ConvNeXt Tiny network with complementary traditional radiomics features, and integrated them with key clinical‑molecular indicators (age, IDH, MGMT) into a visual, clinically applicable nomogram. 2 Materials and methods 2.1 Patients In this retrospective, multicenter study, we consecutively enrolled patients with glioma from two independent centers: Training set: from May 2013 to January 2025, patients were enrolled from the XXX Hospital (Center A). Test set: from May 2016 to January 2025, patients were enrolled from the XXX Hospital (Center B). Patients met the same inclusion and exclusion criteria (Supplementary Fig. 1). The inclusion criteria were as follows: ( 1 ) postoperative pathological confirmation of diffuse glioma, with complete information on WHO grade, IDH mutation status, and MGMT promoter methylation status; ( 2 ) receiving standard treatment, including maximal safe range surgical resection, postoperative concurrent chemoradiotherapy, and at least 6 cycles of temozolomide adjuvant chemotherapy; ( 3 ) new or expanded contrast-enhanced lesions were observed in follow-up MRI; ( 4 ) the suspicious lesion had a clear final classification, which was confirmed by secondary surgical pathology or strict imaging and clinical follow-up. The exclusion criteria were as follows: ( 1 ) receiving radiotherapy, chemotherapy, or other anti-tumor treatments before surgery; ( 2 ) incomplete follow-up MRI image sequences or severe artifacts; and ( 3 ) age less than 18 years. This study was approved by the Ethics Committee of the XXX Hospital of XXX (Ethics Number: XXX), and informed consent was waived. 2.2 Gold Standard and Grouping TP was defined as either histopathologically confirmed recurrence from a second surgery or consecutive MRI‑based progression (enlarging enhancement with increasing edema, mass effect, and clinical decline). PsP was defined as either histopathological evidence of treatment‑related changes without viable tumor cells or MRI follow‑up showing stable/regressing enhancement with reduced edema/mass effect and stable/improved clinical status( 14 ). The study workflow is summarized in Fig. 1 . 2.3 MRI The scanning sequences included axial T1-weighted imaging (T1WI), axial T2-weighted imaging (T2WI), axial diffusion-weighted imaging (DWI), T2 fluid-attenuated inversion recovery (T2-FLAIR) sequence. Before the contrast-enhanced scan, a pre-contrast T1WI was conducted as a mask. Subsequently, gadopentetate dimeglumine (Gd-DTPA) was injected via the cubital vein at a flow rate of 2–3 mL/s, with a dosage of 0.2 mmol/kg body weight. After injection, the axial T1WI sequence was repeated to obtain T1CE images. The imaging parameters of the scanners used at each center are detailed in Supplementary Table S1 . 2.4 Region of Interest (ROI) delineation All ROIs in this study were delineated during the first follow-up MRI scan, which was defined as the time point when a new or enlarged contrast-enhanced lesion was first detected and suspected to be TP or PsP. Before delineating the region of interest, the MRI non-uniformity bias was corrected using the N4 toolbox ( https://github.com/ANTsX/ANTs/wiki/N4BiasFieldCorrection ). Subsequently, using the statistical parametric mapping (SPM12) tool package, the T2WI, T2-FLAIR, and ADC were registered to T1CE image. Then, the images were resampled to a pixel size of 1 mm×1 mm×5 mm using cubic spline interpolation. Next, we used the following formula to conduct intensity normalization: $$\:{I}_{normalized}=100\frac{I-\stackrel{-}{I}}{std\left(I\right)}$$ Radiologist A, who had 5 years of experience in diagnosing MR images of the nervous system, used the ITK-SNAP software (version: 3.4.0) to perform ROI segmentation on the images. T1CE images were selected as the images for tumor segmentation based on a previous study, and segmentation was performed on all tumor layer images( 15 ). One month after the initial segmentation, radiologist A, along with another radiologist B with more than 5 years of experience in neuroimaging diagnosis, independently segmented of 30 randomly assigned patients. 2.5 Training of Deep Learning Models and Feature Extraction This study implemented a ConvNeXt Tiny‑based deep learning model using PyTorch and the Timm library ( https://github.com/huggingface/pytorch-image-models , Supplementary Figure S2)( 16 , 17 ). Model training and feature extraction were performed independently for each sequence. The model initialized the ConvNeXt Tiny network with pre-trained weights from the ImageNet dataset. Next, the model was fine-tuned with the TP or PsP status of patients. All input images and their segmentation masks were uniformly scaled to a resolution of 224×224 pixels. During the training phase, random horizontal flipping and rotation transformations were employed for data augmentation to enhance model robustness and prevent overfitting. The training used cross-entropy as the loss function and employed the AdamW optimizer for optimization, with an initial learning rate of 0.001 and a weight decay strategy. During training, a cosine annealing scheduler was adopted to dynamically adjust the learning rate, with a batch size of 64. Each training level ran for at least 200 cycles and applied the early stopping mechanism. After model training, images in the training and test sets were input into this feature extractor to obtain high-dimensional deep learning features. All experiments in this study were completed on a computing platform equipped with a NVIDIA GeForce RTX 3090 GPU, an Intel Core i7-12700F CPU, and 64 GB of memory. 2.6 Feature Extraction of Radiomics Traditional radiomics features were extracted from the delineated regions using a standardized radiomics pipeline. A detailed description of the feature classes, preprocessing filters, and extraction parameters is provided in the Supplementary Methods S1. Feature Selection and Feature Signature Construction All features were z‑score normalized. Selection proceeded in three steps: retaining features with intra‑ and inter‑class correlation coefficients (ICC) > 0.75; removing highly correlated pairs (Spearman’s |ρ| > 0.9); applying LASSO for final selection. Three machine learning models—Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Machine (SVM)—were trained and optimized using five‑fold cross‑validation. 2.7 Statistical Analysis Statistical analysis was conducted using R software (version 3.6.3, http://www.rproject.org ). Independent sample t-tests or Mann-Whitney U tests were used to compare two groups of continuous variables. Fisher's exact probability method or chi-square test was applied for categorical variables. A two-sided P value less than 0.05 was used as the criterion for significant differences. This study constructed deep learning signatures using logistic regression. For each patient, the average of all two‑dimensional layer features from each image set was computed as the final signature value. Independent predictors were identified through stepwise multivariate logistic regression, with the Akaike information criterion (AIC) used as the stopping rule. Based on the resulting regression coefficients, a deep learning‑radiomics nomogram (DLRN) was established. For comparison, three additional models were established: a clinical model, a radiomics model, and a deep learning model. The diagnostic performance of DLRN was compared with evaluations by two neuroimaging radiologists (C and D, each with > 5 years of experience). Blinded to patient identity, both radiologists independently reviewed all test‑set cases using the same multi‑parameter MRI sequences and classified each as TP or PsP. Model discriminative ability was assessed using the receiver operating characteristic (ROC) curve and area under the curve (AUC), with comparisons performed via the Delong test. Improvement in predictive performance was quantified by the net reclassification index (NRI). Calibration was evaluated visually with calibration curves and statistically with the Hosmer–Lemeshow test. Clinical utility was measured using decision curve analysis (DCA). 3 Results 3.1 Baseline Balance Analysis Patients in the training and test sets exhibited good balance in terms of baseline clinical, pathological, and imaging characteristics (Table 1 ). Particularly, the gold standard diagnostic criterion for the study endpoint, the proportion of true tumor progression diagnosed based on pathological examination shows no significant difference ( P = 0.865). Table 1 Clinical characteristics of patients in different sets Characteristic Training set Test set P Sample size 201 90 Age (year, mean ± SD) 50.9 ± 10.7 52.0 ± 10.3 0.424 Gender (%) 0.822 Male 124 (61.1%) 53 (58.9%) Female 79 (38.9%) 37 (41.1%) Grade (%) 0.809 II 40 (19.7%) 16 (17.8%) III 38 (18.7%) 15 (16.7%) IV 125 (61.6%) 59 (65.6%) IDH (%) 0.539 Negative 102 (50.2%) 41 (45.6%) Positive 101 (49.8%) 49 (54.4%) MGMT (%) 0.718 Negative 119 (58.6%) 50 (55.6%) Positive 84 (41.4%) 40 (44.4%) Read 1 diagnosis (%) > 0.999 PsP 100 (49.3%) 44 (48.9%) TP 103 (50.7%) 46 (51.1%) Read 2 diagnosis (%) 0333 PsP 86 (42.4%) 32 (35.6%) TP 117 (57.6%) 58 (64.4%) Histopathological diagnosis (%) 0.865 PsP 60 (29.6%) 25 (27.8%) TP 143 (70.4%) 65 (72.2%) Note: Independent samples t-test was applied in continuous variables. Chi-Squared test or Fisher’s exact test was applied in categorical variables. Bold type indicates statistically significant difference. IDH = isocitrate dehydrogenase; MGMT = O⁶-methylguanine-DNA methyltransferase; PsP = pseudoprogression; TP = true progression; SD: standard deviation 3.2 Selection of Clinical Characteristics Three preoperative characteristics, namely age, IDH, and MGMT, were used to establish a clinical model (AIC = 235.1). The AUC of the clinical model was 0.688 [95% confidence interval (95% CI): 0.621–0.749] in the training set and 0.630 (95% CI: 0.533–0.733) in the test set (Supplementary Table S2, Fig. 2 ). The odds ratios (OR) of each component variable are shown in Supplementary Table S3. 3.3 Construction of Imaging Signatures After excluding non-repeatable and redundant features, 31 remaining features were left (Supplementary Table S4). Classifier selection based on test-set accuracy identified an SVM with an RBF kernel as optimal for constructing radiomics signatures (Supplementary Table S5). The radiomics signatures of T1CE and ADC exhibited satisfactory performance in the test set, with an AUC of 0.714 (95% CI: 0.622-0.800) for the former and an AUC of 0.830 (95% CI: 0.744-0.900) for the latter in the test set. However, the AUC of the T2-FLAIR signature in the test set was only 0.562 (95% CI: 0.456–0.667) (Supplementary Table S2). 3.4 Construction of DLRN Model and Its Application in Individualized TP Risk Prediction Based on univariate and multivariate logistic regression analyses, five preoperative indicators — age, IDH status, MGMT status, T1CE signature, and ADC signature — were used to construct the DLRN model (AIC = 148.2, Table 2 ). Next, using the regression coefficients derived from the derivation, a DLRN was established. The nomogram model constructed in this study can calculate a patient's risk of TP through the steps showing in Figs. 3 and 4 . Table 2 Variables and coefficients of the nomogram model Variables Univariable OR (95% CI) Pa Adjusted OR (95% CI) Pb Intercept - 0.410(0.048, 3.515) 0.416 Age (per 1-year increase) 1.018(1.012, 1.024) < 0.001 1.050(1.006, 1.097) 0.026 Grade (per 1-grade increase) 1.774(1.458, 2.157) < 0.001 1.280(0.760, 2.154) 0.353 II Ref. Ref. III 0.825 (0.333, 2.045) 0.678 0.284 (0.076, 1.070) 0.063 IV 1.986 (1.074, 4.260) 0.038 1.200 (0.409, 3.518) 0.740 IDH (wild-type vs. mutante) 1.658(1.108, 2.480) 0.014 0.331(0.134, 0.816) 0.016 MGMT (unmethylated vs. methylated) 1.545(1.030, 2.395) 0.049 0.343(0.139, 0.850) 0.021 T2-FLAIR (per 1-unit increase) 1.781(1.308, 2.426) < 0.001 1.334(0.829, 2.147) 0.235 T1CE (per 1-unit increase) 3.639 (2.464, 5.373) < 0.001 2.740(1.694, 4.431) < 0.001 ADC (per 1-unit increase) 3.256(2.276, 4.657) < 0.001 3.893(2.348, 6.456) < 0.001 Note: IDH = isocitrate dehydrogenase; MGMT = O⁶-methylguanine-DNA methyltransferase; PsP = pseudoprogression; OR = odds ratio; T1CE = T1-weighted contrast-enhanced imaging; ADC = apparent diffusion coefficient; P a := P values from univariate logistic regression analysis; P b =P values from multivariable logistic regression analysis (model adjusted for all variables listed in the table) 3.5 Performance, Clinical Utility, and Radiologist-Assistive Value of the DLRN Diagnostic Model As shown in Supplementary Table S2 and Fig. 2 , the developed diagnostic model DLRN exhibited strong predictive ability in both the training and test sets, with AUC values of 0.907 (95% CI: 0.867–0.946) and 0.873 (95% CI: 0.800-0.933), respectively. These AUC values were higher than those of the clinical model (Delong test, P values all less than 0.001). DLRN was significantly superior to the clinical model, the single T1CE signature, and the single ADC signature in both the training and test sets (Supplementary Table S6). Further model comparison indicated that the constructed nomogram was superior to the manual radiomics model in all datasets, significantly superior to the deep learning feature model in the training set. In addition, the calibration curve of the diagnostic model showed good consistency between the observed results and the predicted results in both datasets (Hosmer-Lemeshow test, P values were 0.966 and 0.713, respectively) (Fig. 5 A). As shown in Fig. 5 B, decision curve analysis revealed that the DLRN model achieves superior net clinical benefit across a wide range of thresholds compared to other models when intervening only on predicted high-risk patients. The DLRN model achieved significantly higher accuracy (85.6%) and AUC (0.873) on the test set compared to the two assessors (accuracy: 63.3% and 68.9%; AUC: 0.660 and 0.686) (Supplementary Table S2). Specifically, the DLRN model indicated the greatest advantage in terms of sensitivity, indicating that it can more effectively identify TP cases and reduce false negative rates ( P < 0.05). In the test set, the DLRN model significantly improved the diagnostic performance of radiologists C (Rc) and D (Rd), with net reclassification improvement (NRI) values of 1.102 (95% CI: 0.685–1.490) and 1.298 (95% CI: 0.943–1.625), respectively (Supplementary Table S7). 4 Discussion This study presents a DLRN model that integrates ConvNeXt Tiny–based deep learning features, radiomics, and clinical indicators to distinguish TP from PsP using conventional MRI. By capturing both imaging patterns and patient molecular background, the model enables accurate early risk stratification, helping to avoid delayed or excessive treatment and offering greater clinical value than visual assessment alone. The manual radiomics model in this study showed limited performance in differentiating TP from PsP (test AUC = 0.617). Similarly, a T1CE-based texture model achieved only moderate accuracy (ACC: 72.78%)( 18 ). Integrating clinical features has been shown to improve performance (AUC = 0.729), suggesting that imaging features alone are insufficient( 19 ). Although radiomics from advanced functional MRI sequences (e.g., DWI, PWI) and multiparametric strategies can provide additional value (AUC up to 0.85)( 20 ), the pure manual radiomics model in this study still exhibited limited predictive efficacy in the test set( 21 ). The deep learning feature model constructed in this study demonstrated excellent discriminatory efficacy, with an AUC of 0.814 in the test set. In recent years, CNN, ViT, and their variants have been widely applied in differentiating glioma progression. For instance, Bacchi et al. achieved preliminary validation based on a multi-sequence CNN model in a small sample (accuracy 0.82), suggesting the potential of multi-sequence fusion( 22 ). Another study used a ViT model on an open dataset to confirm the baseline efficacy of a pure deep learning model (AUC ≈ 0.717)( 23 ). Additionally, a study attempted to use complex architectures, such as CNN-LSTM, to process multiple sequences and achieved degrees of improvement in small samples (AUC 0.81)( 24 ). Compared to the above studies, this study has systematically optimized the following aspects: ( 1 ) Data level: A large multicenter cohort was adopted, thereby enhancing the model’s generalization reliability; ( 2 ) Model architecture: The ConvNeXt Tiny network with superior performance was employed, overcoming the limited representational capacity of traditional CNNs( 17 ); ( 3 ) Strategy level: Deep learning features, traditional radiomics features, and key clinical molecular markers were innovatively integrated, achieving performance improvement based on comprehensive information. This study developed and validated a nomogram model integrating deep learning imaging features, conventional imaging biomarkers, and key clinical molecular indicators, which demonstrated excellent performance in distinguishing true progression from pseudoprogression in gliomas. A recent study indirectly confirms the universal validity and strong potential of the multimodal fusion strategy in neuro-tumor image analysis( 25 ). Using the nomogram proposed in this study, clinicians can directly convert the clinical features of patients, multi-sequence imaging biomarker signatures, and other indicators into specific scores. The individual score and the predicted probability of TP can be obtained by simply adding the scores of all indicators. Through rigorous auxiliary diagnosis experiments, this study proved that with the assistance of the DLRN model, the diagnostic accuracy of both radiologists increased to 80%, and the sensitivity reached 83.1% and 76.9%, respectively. This finding confirms that the DLRN model not only serves as an independent diagnostic tool but also provides an effective clinical decision support system. This study has several limitations. As a retrospective multicenter analysis, variations in scanning parameters and potential selection bias may have influenced performance despite standardized processing. The limited number of centers may restrict robustness, warranting future validation across more institutions, scanners, and protocols. Moreover, the cohort mainly included diagnostically challenging cases biased toward TP, which may limit generalizability to all patients with suspected progression. 5 Conclusions This study developed a ConvNeXt Tiny–based multimodal DLRN model that outperformed single-modal models in distinguishing TP from PsP in gliomas and significantly improved radiologists’ diagnostic accuracy with model assistance. Declarations 7.1 Ethics approval and consent to participate This retrospective study was conducted in accordance with the ethical standards of the Declaration of Helsinki and was approved by the Ethics Review Board of the XXX Hospital of XXX (Ethics Number: XXXXXX). The requirement for written informed consent from participants was waived by the approving ethics committee due to the retrospective nature of the study. 7.2 Consent for publication The need for written informed consent was waived with the confirmation of patient data confidentiality by the institutional Ethics Committee for this retrospective study. 7.3 Availability of data and materials The data that support the findings of this study are not publicly available due to privacy or ethical restrictions, but are available on request from the corresponding author. 7.4 Competing interests The authors declare that they have no competing interests. 7.5 Funding This work was supported by Qinhuangdao S&T Plan Program (grant number: 202301A212) 7.6 Author Contributions T.Z.: Conceptualization, Writing – original draft. L.Y.: Data curation, Writing – original draft. D.Z.: Methodology. J.D.: Funding acquisition. X.L.: Formal analysis. S.W.: Visualization. X.W.: Supervision. Q.S.: Software. D.L.: Writing - Reviewing and Editing. 7.7 Acknowledgments Not applicable 7.8 Clinical trial number Not applicable. References Weller M, Wen PY, Chang SM, Dirven L, Lim M, Monje M, Reifenberger G, Glioma (2024) Nat reviews Disease primers 10(1):33. 10.1038/s41572-024-00516-y Kowalczyk A, Zarychta J, Marszołek A, Zawitkowska J, Lejman M (2024) Chimeric Antigen Receptor T Cell and Chimeric Antigen Receptor NK Cell Therapy in Pediatric and Adult High-Grade Glioma-Recent Advances. Cancers 16(3). 10.3390/cancers16030623 Wen PY, van den Bent M, Youssef G, Cloughesy TF, Ellingson BM, Weller M, Galanis E, Barboriak DP, de Groot J, Gilbert MR, Huang R, Lassman AB, Mehta M, Molinaro AM, Preusser M, Rahman R, Shankar LK, Stupp R, Villanueva-Meyer JE, Wick W, Macdonald DR, Reardon DA, Vogelbaum MA, Chang SM (2023) RANO 2.0: Update to the Response Assessment in Neuro-Oncology Criteria for High- and Low-Grade Gliomas in Adults. J Clin oncology: official J Am Soc Clin Oncol 41(33):5187–5199. 10.1200/jco.23.01059 Hygino da Cruz LC Jr., Rodriguez I, Domingues RC, Gasparetto EL, Sorensen AG (2011) Pseudoprogression and pseudoresponse: imaging challenges in the assessment of posttreatment glioma. AJNR Am J Neuroradiol 32(11):1978–1985. 10.3174/ajnr.A2397 Abdalla G, Hammam A, Anjari M, D'Arco DF, Bisdas DS (2020) Glioma surveillance imaging: current strategies, shortcomings, challenges and outlook. BJR open 2(1):20200009. 10.1259/bjro.20200009 Calmon R, Puget S, Varlet P, Dangouloff-Ros V, Blauwblomme T, Beccaria K, Grevent D, Sainte-Rose C, Castel D, Debily MA, Dufour C, Bolle S, Dhermain F, Saitovitch A, Zilbovicius M, Brunelle F, Grill J, Boddaert N (2018) Cerebral blood flow changes after radiation therapy identifies pseudoprogression in diffuse intrinsic pontine gliomas. Neurooncology 20(7):994–1002. 10.1093/neuonc/nox227 Bogsrud TV, Londalen A, Brandal P, Leske H, Panagopoulos I, Borghammer P, Bach-Gansmo T (2019) 18F-Fluciclovine PET/CT in Suspected Residual or Recurrent High-Grade Glioma. Clin Nucl Med 44(8):605–611. 10.1097/rlu.0000000000002641 Reddy S, Lung T, Muniyappa S, Hadley C, Templeton B, Fritz J, Boulter D, Shah K, Singh R, Zhu S, Matsui JK, Palmer JD (2025) Radiomics and Radiogenomics in Differentiating Progression, Pseudoprogression, and Radiation Necrosis in Gliomas. Biomedicines 13(7). 10.3390/biomedicines13071778 Li J, Xu Q, Fan X, Cheng X, Tao J, Lu H, Lin Q, Zhang J, Qian J (2025) Multi-Sequence MRI radiomics model for discrimination of recurrence and pseudoprogression in gliomas. Eur J Radiol 194:112508. 10.1016/j.ejrad.2025.112508 Mayerhoefer ME, Materka A, Langs G, Häggström I, Szczypiński P, Gibbs P, Cook G (2020) Introduction to Radiomics. Journal of nuclear medicine: official publication. Soc Nuclear Med 61(4):488–495. 10.2967/jnumed.118.222893 Tripathi PC, Bag S (2023) An Attention-Guided CNN Framework for Segmentation and Grading of Glioma Using 3D MRI Scans. IEEE/ACM transactions on computational biology and bioinformatics. ;20(3):1890–1904. 10.1109/tcbb.2022.3220902 Cheng J, Liu J, Kuang H, Wang JA, Fully Automated (2022) Multimodal MRI-Based Multi-Task Learning for Glioma Segmentation and IDH Genotyping. IEEE Trans Med Imaging 41(6):1520–1532. 10.1109/tmi.2022.3142321 Xu C, Peng Y, Zhu W, Chen Z, Li J, Tan W, Zhang Z, Chen X (2022) An automated approach for predicting glioma grade and survival of LGG patients using CNN and radiomics. Front Oncol 12:969907. 10.3389/fonc.2022.969907 Wen PY, Macdonald DR, Reardon DA, Cloughesy TF, Sorensen AG, Galanis E, Degroot J, Wick W, Gilbert MR, Lassman AB, Tsien C, Mikkelsen T, Wong ET, Chamberlain MC, Stupp R, Lamborn KR, Vogelbaum MA, van den Bent MJ, Chang SM (2010) Updated response assessment criteria for high-grade gliomas: response assessment in neuro-oncology working group. J Clin oncology: official J Am Soc Clin Oncol 28(11):1963–1972. 10.1200/jco.2009.26.3541 De Sutter S, Wuts J, Geens W, Vanbinst AM, Duerinck J, Vandemeulebroucke J (2024) Modality redundancy for MRI-based glioblastoma segmentation. Int J Comput Assist Radiol Surg 19(10):2101–2109. 10.1007/s11548-024-03238-4 Nigam S, Dheeraj A, Sachan H, Marwaha S (2025) Automated weed classification using attention-embedded ConvNeXtV2 architecture. Procedia Comput Sci 260:291–299. https://doi.org/10.1016/j.procs.2025.03.204 Liu Z, Mao H, Wu C-Y, Feichtenhofer C, Darrell T, Xie S A ConvNet for the 2020s. arXiv 2022. https://arxiv.org/abs/2201.03545 Sun YZ, Yan LF, Han Y, Nan HY, Xiao G, Tian Q, Pu WH, Li ZY, Wei XC, Wang W, Cui GB (2021) Differentiation of Pseudoprogression from True Progressionin Glioblastoma Patients after Standard Treatment: A Machine Learning Strategy Combinedwith Radiomics Features from T(1)-weighted Contrast-enhanced Imaging. BMC Med Imaging 21(1):17. 10.1186/s12880-020-00545-5 Ari AP, Akkurt BH, Musigmann M, Mammadov O, Blömer DA, Kasap DNG, Henssen D, Nacul NG, Sartoretti E, Sartoretti T, Backhaus P, Thomas C, Stummer W, Heindel W, Mannil M (2022) Pseudoprogression prediction in high grade primary CNS tumors by use of radiomics. Sci Rep 12(1):5915. 10.1038/s41598-022-09945-9 Kim JY, Park JE, Jo Y, Shim WH, Nam SJ, Kim JH, Yoo RE, Choi SH, Kim HS (2019) Incorporating diffusion- and perfusion-weighted MRI into a radiomics model improves diagnostic performance for pseudoprogression in glioblastoma patients. Neurooncology 21(3):404–414. 10.1093/neuonc/noy133 Sala E, Mema E, Himoto Y, Veeraraghavan H, Brenton JD, Snyder A, Weigelt B, Vargas HA (2017) Unravelling tumour heterogeneity using next-generation imaging: radiomics, radiogenomics, and habitat imaging. Clin Radiol 72(1):3–10. 10.1016/j.crad.2016.09.013 Bacchi S, Zerner T, Dongas J, Asahina AT, Abou-Hamden A, Otto S, Oakden-Rayner L, Patel S (2019) Deep learning in the detection of high-grade glioma recurrence using multiple MRI sequences: A pilot study. J Clin neuroscience: official J Neurosurgical Soc Australasia 70:11–13. 10.1016/j.jocn.2019.10.003 Gomaa A, Huang Y, Stephan P, Breininger K, Frey B, Dörfler A, Schnell O, Delev D, Coras R, Donaubauer AJ, Schmitter C, Stritzelberger J, Semrau S, Maier A, Bayer S, Schönecker S, Heiland DH, Hau P, Gaipl US, Bert C, Fietkau R, Schmidt MA, Putz F (2025) A self-supervised multimodal deep learning approach to differentiate post-radiotherapy progression from pseudoprogression in glioblastoma. Sci Rep 15(1):17133. 10.1038/s41598-025-02026-7 Lee J, Wang N, Turk S, Mohammed S, Lobo R, Kim J, Liao E, Camelo-Piragua S, Kim M, Junck L, Bapuraj J, Srinivasan A, Rao A (2020) Discriminating pseudoprogression and true progression in diffuse infiltrating glioma using multi-parametric MRI data through deep learning. Sci Rep 10(1):20331. 10.1038/s41598-020-77389-0 Bijari S, Rezaeijo SM, Sayfollahi S, Rahimnezhad A, Heydarheydari S (2025) Development and validation of a robust MRI-based nomogram incorporating radiomics and deep features for preoperative glioma grading: a multi-center study. Quant imaging Med Surg 15(2):1125–1138. 10.21037/qims-24-1543 Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx 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-8643558","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":578897009,"identity":"60a9e021-273c-4258-9880-0fdf2a6747dc","order_by":0,"name":"Tao Zheng","email":"","orcid":"","institution":"The First Hospital of Qinhuangdao","correspondingAuthor":false,"prefix":"","firstName":"Tao","middleName":"","lastName":"Zheng","suffix":""},{"id":578897010,"identity":"93824eab-bb4b-4d09-91b0-ab7a0d8a907c","order_by":1,"name":"Linsha Yang","email":"","orcid":"","institution":"The First Hospital of Qinhuangdao","correspondingAuthor":false,"prefix":"","firstName":"Linsha","middleName":"","lastName":"Yang","suffix":""},{"id":578897011,"identity":"ef5f2a88-a922-4571-9a50-9bf45cfff6b9","order_by":2,"name":"Duo Zhang","email":"","orcid":"","institution":"Baoding No.1 Hospital","correspondingAuthor":false,"prefix":"","firstName":"Duo","middleName":"","lastName":"Zhang","suffix":""},{"id":578897012,"identity":"dffa5695-beef-463c-b706-166ac86956e1","order_by":3,"name":"Juan Du","email":"","orcid":"","institution":"The First Hospital of Qinhuangdao","correspondingAuthor":false,"prefix":"","firstName":"Juan","middleName":"","lastName":"Du","suffix":""},{"id":578897013,"identity":"9f3af609-1c69-40e2-a98f-3e7b49be9ed0","order_by":4,"name":"Xin Liang","email":"","orcid":"","institution":"The First Hospital of Qinhuangdao","correspondingAuthor":false,"prefix":"","firstName":"Xin","middleName":"","lastName":"Liang","suffix":""},{"id":578897014,"identity":"94149ea9-7fa5-43a3-bc55-51dd91ef61a9","order_by":5,"name":"Shuo Wu","email":"","orcid":"","institution":"The First Hospital of Qinhuangdao","correspondingAuthor":false,"prefix":"","firstName":"Shuo","middleName":"","lastName":"Wu","suffix":""},{"id":578897015,"identity":"f52b58c0-604e-42c3-9dc7-78f67d3631f4","order_by":6,"name":"Xiaohan Wang","email":"","orcid":"","institution":"The First Hospital of Qinhuangdao","correspondingAuthor":false,"prefix":"","firstName":"Xiaohan","middleName":"","lastName":"Wang","suffix":""},{"id":578897016,"identity":"9582cc47-416a-48ef-a83f-ce049fa0f0b2","order_by":7,"name":"Qinglei Shi","email":"","orcid":"","institution":"Siemens Ltd","correspondingAuthor":false,"prefix":"","firstName":"Qinglei","middleName":"","lastName":"Shi","suffix":""},{"id":578897017,"identity":"80d818a5-96b4-488c-9e86-63e6e6904c0d","order_by":8,"name":"Defeng Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+0lEQVRIie3PPWrDMBTAcYkH9iJ7fsIhuYKClgyFXiUmkCn0gy4ZSlEwpEvoXig5RmYFgbPoAN5S1xcIdOlHhgpPnSyPheqPeCB4P4QICYX+Ypq+ujlhJF6t6tMShyM/AeEmMsKMkWgncqx6Enfmc0zWy5xoj0gPBW1uvnAg9EIg3+KUKqjfqg7CrQH5/IRMaCvE7Q6vYhJJueggorous2TjyH4jpnyHd1SxKOskxyb+bolxDyUvmCvtIxVEwD4cKaNcJaoH4XYGmdtk3IIBLFGOC89f0oOh7+x8cZke68fP0/3DcBQXddNF2uj69w18623nXluhUCj0X/sB8GVLoPKJaMUAAAAASUVORK5CYII=","orcid":"","institution":"The First Hospital of Qinhuangdao","correspondingAuthor":true,"prefix":"","firstName":"Defeng","middleName":"","lastName":"Liu","suffix":""}],"badges":[],"createdAt":"2026-01-20 00:53:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8643558/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8643558/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101274415,"identity":"dc9ab3b4-2a7d-4b7f-a35b-e108450c7c7c","added_by":"auto","created_at":"2026-01-28 03:10:22","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":526686,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eWorkflow of the study. \u003c/strong\u003e(A) Delineation of two-dimensional regions of interest (2D ROIs) was performed on post-treatment T2-weighted images, T2-weighted fluid-attenuated inversion recovery (T2-FLAIR) images, T1-weighted contrast-enhanced (T1CE) images, and apparent diffusion coefficient (ADC) maps. The segmentation encompassed the entire enhancing lesion. (B) Feature extraction was conducted using conventional radiomics algorithms and a pre-trained ConvNeXt Tiny convolutional neural network (CNN) to obtain handcrafted radiomics features and deep learning (DL) features from all four MRI sequences. (C) Feature selection was performed in multiple steps, including evaluation of interclass/intraclass correlation coefficient (ICC), Spearman rank correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression with 5-fold cross-validation. (D) Multivariate logistic regression with backward stepwise selection based on Akaike's information criterion was employed to integrate significant clinical factors (age, IDH status, MGMT status) with selected radiomics and DL features, ultimately constructing a comprehensive deep learning radiomics nomogram (DLRN) for discriminating between true progression and pseudoprogression.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8643558/v1/21e81a5e797c9d7e21faf1ad.png"},{"id":101274443,"identity":"91d3b45d-1c3e-4ce3-8be3-28598773c782","added_by":"auto","created_at":"2026-01-28 03:10:28","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":295030,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReceiver operating characteristic (ROC) curves of different signatures/models, along with the expert points of two radiologists with and without the assistance of the deep learning radiomics nomogram (DLRN), in the training set (A) and test set (B).\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8643558/v1/520bd846d1a22913529b65ef.png"},{"id":101274456,"identity":"57fdbedc-0cb1-4e6f-9aa6-8ae1146e04ea","added_by":"auto","created_at":"2026-01-28 03:10:34","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":444173,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExamples of individual risk prediction for tumor progression (TP) using the nomogram. \u003c/strong\u003eA 52-year-old female patient with WHO grade 4 glioblastoma with pathologically confirmed as TP. Firstly, based on the specific indicators of the patient, a vertical line was drawn on the \"Points\" scale at the top of the nomogram to obtain the corresponding score. These indicators included age (red line), IDH status (yellow), MGMT status (blue), T1CE signature (green), and ADC signature (purple). Then, the corresponding scores of all variables were summed to obtain the \"Total Points\" total score. Finally, a black vertical line was drawn at the total score position on the \"Total Points\" axis. The intersection with the bottom the \"Risk of TP\" axis was the individualized prediction risk value of the patient. This is an example of a typical case is as follows: The case was a 52-year-old female patient with WHO grade 4 glioblastoma, wild-type IDH, and non-methylated MGMT. The corresponding scores for each feature were 15, 10, 9, 39, and 43, with a total score of 116. The predicted TP risk was greater than 90%.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8643558/v1/f877409f24b4185c58939ed7.png"},{"id":101274479,"identity":"6e97ae7b-e9a1-4faa-992c-4044c1a14c8c","added_by":"auto","created_at":"2026-01-28 03:10:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":486247,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExamples of individual risk prediction for tumor progression (TP) using the nomogram. \u003c/strong\u003eA 62-year-old male patient with WHO grade 3 glioma with pathologically confirmed as pseudoprogression (PsP). Firstly, based on the specific indicators of the patient, a vertical line was drawn on the \"Points\" scale at the top of the nomogram to obtain the corresponding score. These indicators included age (red line), IDH status (yellow), MGMT status (blue), T1CE signature (green), and ADC signature (purple). Then, the corresponding scores of all variables were summed to obtain the \"Total Points\" total score. Finally, a black vertical line was drawn at the total score position on the \"Total Points\" axis. The intersection with the bottom the \"Risk of TP\" axis was the individualized prediction risk value of the patient. This case was a 62-year-old male patient with a WHO grade 3 glioma, mutant IDH, and methylated MGMT. The corresponding scores for each feature were 18, 0, 0, 26, and 33, with a total score of 77. The predicted TP risk was lower than 20%.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8643558/v1/7e1a8db18b3c8b89b4c30047.png"},{"id":101274493,"identity":"a6939622-dd6a-44e5-a8eb-0184a03458ff","added_by":"auto","created_at":"2026-01-28 03:10:44","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":272101,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCalibration curves and decision curve analysis (DCA). \u003c/strong\u003e(A) Calibration curves of deep learning radiomics nomogram in both sets. (B) DCA for deep learning radiomics nomogram (DLRN), T1CE signature, ADC signature and clinical model.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8643558/v1/41578afb886e4fd0ac9b2ce5.png"},{"id":101274509,"identity":"402444c2-ab9a-4e12-9fde-dbea748edcb6","added_by":"auto","created_at":"2026-01-28 03:10:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3111600,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8643558/v1/bfcd3557-01b6-45cb-aa18-c73fa1998b75.pdf"},{"id":101274449,"identity":"799eab09-33af-4ad1-8327-62efda83e10b","added_by":"auto","created_at":"2026-01-28 03:10:31","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":236292,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-8643558/v1/c517310904a03c8352afff2e.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrating ConvNeXt Tiny and Radiomics in a Nomogram to Differentiate True Progression from Pseudoprogression in Glioma","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eGlioma is the most common primary malignant tumor in the adult central nervous system, accounting for approximately 80% of all malignant brain tumors. For newly diagnosed high-grade gliomas, the standard treatment includes surgical resection in the maximum safe range, followed by concurrent chemoradiotherapy and adjuvant chemotherapy based on temozolomide(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Approximately 20%-30% of patients will have new or expanded enhanced lesions that meet the imaging \"progression\" criteria in the initial follow-up MRI after treatment(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Some of these lesions represent true tumor progression (TP) and require active treatments, such as reoperation and changes in chemotherapy regimen(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). However, another part is pseudo-progression (PsP), an excessive local tissue response confined to the tumor area subsequent to effective treatment(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eCurrently, the clinical differentiation between TP and PsP in glioma primarily relies on imaging follow-up. However, in the early stage after treatment, both TP and PsP can present similar abnormal enhancement foci and peripheral edema on conventional MRI (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). Arterial spin-labeled perfusion imaging revealed that the increase in cerebral blood flow was more significant in patients with signs of PsP(\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). Positron emission computed tomography (PET-CT) scans also showed that the uptake of specific tracers in patients with recurrent high-grade glioma is significantly higher than that in PsP lesions(\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). Despite offering supplementary insights into tumor angiogenesis, cellularity, and metabolism, these advanced functional imaging techniques are limited by non-standardized protocols, high complexity and cost, and restricted routine clinical applicability.\u003c/p\u003e \u003cp\u003eSeveral studies have confirmed that radiomics performs well in differentiating TP from PsP(\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). However, traditional radiomics methods heavily rely on manual delineation of regions of interest, which is not only time-consuming but also introduces inter-observer variability(\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). Deep learning technology can automatically extract complex and deep features from raw images through an end-to-end learning approach, overcoming the limitations of manual design and demonstrating stronger feature expression and generalization capabilities. Currently, various advanced architectures, including convolutional neural networks and vision transformers, have achieved significant progress in automatic segmentation of gliomas, non-invasive prediction of isocitrate dehydrogenase (IDH) mutation status, tumor grading, and prognosis assessment(\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis study aimed to develop and validate a multimodal nomogram to noninvasively differentiate TP from PsP in glioma. Innovatively, we combined deep features extracted automatically by the ConvNeXt Tiny network with complementary traditional radiomics features, and integrated them with key clinical‑molecular indicators (age, IDH, MGMT) into a visual, clinically applicable nomogram.\u003c/p\u003e"},{"header":"2 Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Patients\u003c/h2\u003e \u003cp\u003eIn this retrospective, multicenter study, we consecutively enrolled patients with glioma from two independent centers: Training set: from May 2013 to January 2025, patients were enrolled from the XXX Hospital (Center A). Test set: from May 2016 to January 2025, patients were enrolled from the XXX Hospital (Center B).\u003c/p\u003e \u003cp\u003ePatients met the same inclusion and exclusion criteria (Supplementary Fig.\u0026nbsp;1). The inclusion criteria were as follows: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) postoperative pathological confirmation of diffuse glioma, with complete information on WHO grade, IDH mutation status, and MGMT promoter methylation status; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) receiving standard treatment, including maximal safe range surgical resection, postoperative concurrent chemoradiotherapy, and at least 6 cycles of temozolomide adjuvant chemotherapy; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) new or expanded contrast-enhanced lesions were observed in follow-up MRI; (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) the suspicious lesion had a clear final classification, which was confirmed by secondary surgical pathology or strict imaging and clinical follow-up. The exclusion criteria were as follows: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) receiving radiotherapy, chemotherapy, or other anti-tumor treatments before surgery; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) incomplete follow-up MRI image sequences or severe artifacts; and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) age less than 18 years. This study was approved by the Ethics Committee of the XXX Hospital of XXX (Ethics Number: XXX), and informed consent was waived.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Gold Standard and Grouping\u003c/h2\u003e \u003cp\u003eTP was defined as either histopathologically confirmed recurrence from a second surgery or consecutive MRI‑based progression (enlarging enhancement with increasing edema, mass effect, and clinical decline). PsP was defined as either histopathological evidence of treatment‑related changes without viable tumor cells or MRI follow‑up showing stable/regressing enhancement with reduced edema/mass effect and stable/improved clinical status(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). The study workflow is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 MRI\u003c/h2\u003e \u003cp\u003eThe scanning sequences included axial T1-weighted imaging (T1WI), axial T2-weighted imaging (T2WI), axial diffusion-weighted imaging (DWI), T2 fluid-attenuated inversion recovery (T2-FLAIR) sequence. Before the contrast-enhanced scan, a pre-contrast T1WI was conducted as a mask. Subsequently, gadopentetate dimeglumine (Gd-DTPA) was injected via the cubital vein at a flow rate of 2\u0026ndash;3 mL/s, with a dosage of 0.2 mmol/kg body weight. After injection, the axial T1WI sequence was repeated to obtain T1CE images. The imaging parameters of the scanners used at each center are detailed in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Region of Interest (ROI) delineation\u003c/h2\u003e \u003cp\u003eAll ROIs in this study were delineated during the first follow-up MRI scan, which was defined as the time point when a new or enlarged contrast-enhanced lesion was first detected and suspected to be TP or PsP.\u003c/p\u003e \u003cp\u003eBefore delineating the region of interest, the MRI non-uniformity bias was corrected using the N4 toolbox (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/ANTsX/ANTs/wiki/N4BiasFieldCorrection\u003c/span\u003e\u003cspan address=\"https://github.com/ANTsX/ANTs/wiki/N4BiasFieldCorrection\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Subsequently, using the statistical parametric mapping (SPM12) tool package, the T2WI, T2-FLAIR, and ADC were registered to T1CE image. Then, the images were resampled to a pixel size of 1 mm\u0026times;1 mm\u0026times;5 mm using cubic spline interpolation. Next, we used the following formula to conduct intensity normalization:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{I}_{normalized}=100\\frac{I-\\stackrel{-}{I}}{std\\left(I\\right)}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eRadiologist A, who had 5 years of experience in diagnosing MR images of the nervous system, used the ITK-SNAP software (version: 3.4.0) to perform ROI segmentation on the images. T1CE images were selected as the images for tumor segmentation based on a previous study, and segmentation was performed on all tumor layer images(\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOne month after the initial segmentation, radiologist A, along with another radiologist B with more than 5 years of experience in neuroimaging diagnosis, independently segmented of 30 randomly assigned patients.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Training of Deep Learning Models and Feature Extraction\u003c/h2\u003e \u003cp\u003eThis study implemented a ConvNeXt Tiny‑based deep learning model using PyTorch and the Timm library (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/huggingface/pytorch-image-models\u003c/span\u003e\u003cspan address=\"https://github.com/huggingface/pytorch-image-models\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e, Supplementary Figure S2)(\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). Model training and feature extraction were performed independently for each sequence.\u003c/p\u003e \u003cp\u003eThe model initialized the ConvNeXt Tiny network with pre-trained weights from the ImageNet dataset. Next, the model was fine-tuned with the TP or PsP status of patients. All input images and their segmentation masks were uniformly scaled to a resolution of 224\u0026times;224 pixels. During the training phase, random horizontal flipping and rotation transformations were employed for data augmentation to enhance model robustness and prevent overfitting.\u003c/p\u003e \u003cp\u003eThe training used cross-entropy as the loss function and employed the AdamW optimizer for optimization, with an initial learning rate of 0.001 and a weight decay strategy. During training, a cosine annealing scheduler was adopted to dynamically adjust the learning rate, with a batch size of 64. Each training level ran for at least 200 cycles and applied the early stopping mechanism. After model training, images in the training and test sets were input into this feature extractor to obtain high-dimensional deep learning features. All experiments in this study were completed on a computing platform equipped with a NVIDIA GeForce RTX 3090 GPU, an Intel Core i7-12700F CPU, and 64 GB of memory.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Feature Extraction of Radiomics\u003c/h2\u003e \u003cp\u003eTraditional radiomics features were extracted from the delineated regions using a standardized radiomics pipeline. A detailed description of the feature classes, preprocessing filters, and extraction parameters is provided in the Supplementary Methods S1.\u003c/p\u003e \u003cp\u003e \u003cem\u003eFeature Selection and Feature Signature Construction\u003c/em\u003e \u003c/p\u003e \u003cp\u003eAll features were z‑score normalized. Selection proceeded in three steps: retaining features with intra‑ and inter‑class correlation coefficients (ICC)\u0026thinsp;\u0026gt;\u0026thinsp;0.75; removing highly correlated pairs (Spearman\u0026rsquo;s |ρ| \u0026gt; 0.9); applying LASSO for final selection. Three machine learning models\u0026mdash;Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Machine (SVM)\u0026mdash;were trained and optimized using five‑fold cross‑validation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Statistical Analysis\u003c/h2\u003e \u003cp\u003eStatistical analysis was conducted using R software (version 3.6.3, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.rproject.org\u003c/span\u003e\u003cspan address=\"http://www.rproject.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Independent sample t-tests or Mann-Whitney U tests were used to compare two groups of continuous variables. Fisher's exact probability method or chi-square test was applied for categorical variables. A two-sided \u003cem\u003eP\u003c/em\u003e value less than 0.05 was used as the criterion for significant differences.\u003c/p\u003e \u003cp\u003eThis study constructed deep learning signatures using logistic regression. For each patient, the average of all two‑dimensional layer features from each image set was computed as the final signature value. Independent predictors were identified through stepwise multivariate logistic regression, with the Akaike information criterion (AIC) used as the stopping rule. Based on the resulting regression coefficients, a deep learning‑radiomics nomogram (DLRN) was established. For comparison, three additional models were established: a clinical model, a radiomics model, and a deep learning model. The diagnostic performance of DLRN was compared with evaluations by two neuroimaging radiologists (C and D, each with \u0026gt;\u0026thinsp;5 years of experience). Blinded to patient identity, both radiologists independently reviewed all test‑set cases using the same multi‑parameter MRI sequences and classified each as TP or PsP.\u003c/p\u003e \u003cp\u003eModel discriminative ability was assessed using the receiver operating characteristic (ROC) curve and area under the curve (AUC), with comparisons performed via the Delong test. Improvement in predictive performance was quantified by the net reclassification index (NRI). Calibration was evaluated visually with calibration curves and statistically with the Hosmer\u0026ndash;Lemeshow test. Clinical utility was measured using decision curve analysis (DCA).\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Baseline Balance Analysis\u003c/h2\u003e \u003cp\u003ePatients in the training and test sets exhibited good balance in terms of baseline clinical, pathological, and imaging characteristics (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Particularly, the gold standard diagnostic criterion for the study endpoint, the proportion of true tumor progression diagnosed based on pathological examination shows no significant difference (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.865).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClinical characteristics of patients in different sets\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTest set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSample size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e201\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (year, mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e50.9\u0026thinsp;\u0026plusmn;\u0026thinsp;10.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e52.0\u0026thinsp;\u0026plusmn;\u0026thinsp;10.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.424\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGender (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.822\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e124 (61.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e53 (58.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e79 (38.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e37 (41.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrade (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.809\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e40 (19.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16 (17.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38 (18.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15 (16.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e125 (61.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e59 (65.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIDH (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.539\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e102 (50.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e41 (45.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e101 (49.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49 (54.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMGMT (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.718\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e119 (58.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e50 (55.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e84 (41.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e40 (44.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRead 1 diagnosis (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;0.999\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePsP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e100 (49.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e44 (48.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e103 (50.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46 (51.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRead 2 diagnosis (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0333\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePsP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e86 (42.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e32 (35.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e117 (57.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e58 (64.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHistopathological diagnosis (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.865\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePsP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e60 (29.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25 (27.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e143 (70.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e65 (72.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003eNote: Independent samples t-test was applied in continuous variables. Chi-Squared test or Fisher\u0026rsquo;s exact test was applied in categorical variables. Bold type indicates statistically significant difference. IDH\u0026thinsp;=\u0026thinsp;isocitrate dehydrogenase; MGMT\u0026thinsp;=\u0026thinsp;O⁶-methylguanine-DNA methyltransferase; PsP\u0026thinsp;=\u0026thinsp;pseudoprogression; TP\u0026thinsp;=\u0026thinsp;true progression; SD: standard deviation\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Selection of Clinical Characteristics\u003c/h2\u003e \u003cp\u003eThree preoperative characteristics, namely age, IDH, and MGMT, were used to establish a clinical model (AIC\u0026thinsp;=\u0026thinsp;235.1). The AUC of the clinical model was 0.688 [95% confidence interval (95% CI): 0.621\u0026ndash;0.749] in the training set and 0.630 (95% CI: 0.533\u0026ndash;0.733) in the test set (Supplementary Table S2, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The odds ratios (OR) of each component variable are shown in Supplementary Table S3.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Construction of Imaging Signatures\u003c/h2\u003e \u003cp\u003eAfter excluding non-repeatable and redundant features, 31 remaining features were left (Supplementary Table S4). Classifier selection based on test-set accuracy identified an SVM with an RBF kernel as optimal for constructing radiomics signatures (Supplementary Table S5).\u003c/p\u003e \u003cp\u003eThe radiomics signatures of T1CE and ADC exhibited satisfactory performance in the test set, with an AUC of 0.714 (95% CI: 0.622-0.800) for the former and an AUC of 0.830 (95% CI: 0.744-0.900) for the latter in the test set. However, the AUC of the T2-FLAIR signature in the test set was only 0.562 (95% CI: 0.456\u0026ndash;0.667) (Supplementary Table S2).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Construction of DLRN Model and Its Application in Individualized TP Risk Prediction\u003c/h2\u003e \u003cp\u003eBased on univariate and multivariate logistic regression analyses, five preoperative indicators \u0026mdash; age, IDH status, MGMT status, T1CE signature, and ADC signature \u0026mdash; were used to construct the DLRN model (AIC\u0026thinsp;=\u0026thinsp;148.2, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Next, using the regression coefficients derived from the derivation, a DLRN was established. The nomogram model constructed in this study can calculate a patient's risk of TP through the steps showing in Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eVariables and coefficients of the nomogram model\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariables\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnivariable OR (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003ePa\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAdjusted OR (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003ePb\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIntercept\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.410(0.048, 3.515)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.416\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (per 1-year increase)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.018(1.012, 1.024)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.050(1.006, 1.097)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.026\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrade (per 1-grade increase)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.774(1.458, 2.157)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.280(0.760, 2.154)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.353\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRef.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRef.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.825 (0.333, 2.045)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.678\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.284 (0.076, 1.070)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.063\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.986 (1.074, 4.260)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.038\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.200 (0.409, 3.518)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.740\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIDH (wild-type vs. mutante)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.658(1.108, 2.480)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.014\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.331(0.134, 0.816)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.016\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMGMT (unmethylated vs. methylated)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.545(1.030, 2.395)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.049\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.343(0.139, 0.850)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.021\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT2-FLAIR (per 1-unit increase)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.781(1.308, 2.426)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.334(0.829, 2.147)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.235\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT1CE (per 1-unit increase)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.639 (2.464, 5.373)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.740(1.694, 4.431)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eADC (per 1-unit increase)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.256(2.276, 4.657)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.893(2.348, 6.456)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eNote: IDH\u0026thinsp;=\u0026thinsp;isocitrate dehydrogenase; MGMT\u0026thinsp;=\u0026thinsp;O⁶-methylguanine-DNA methyltransferase; PsP\u0026thinsp;=\u0026thinsp;pseudoprogression; OR\u0026thinsp;=\u0026thinsp;odds ratio; T1CE\u0026thinsp;=\u0026thinsp;T1-weighted contrast-enhanced imaging; ADC\u0026thinsp;=\u0026thinsp;apparent diffusion coefficient; \u003cem\u003eP\u003c/em\u003e\u003csub\u003e\u003cem\u003ea\u003c/em\u003e\u003c/sub\u003e:=\u003cem\u003eP\u003c/em\u003e values from univariate logistic regression analysis; \u003cem\u003eP\u003c/em\u003e\u003csub\u003e\u003cem\u003eb\u003c/em\u003e\u003c/sub\u003e\u003cem\u003e=P\u003c/em\u003e values from multivariable logistic regression analysis (model adjusted for all variables listed in the table)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Performance, Clinical Utility, and Radiologist-Assistive Value of the DLRN Diagnostic Model\u003c/h2\u003e \u003cp\u003eAs shown in Supplementary Table S2 and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the developed diagnostic model DLRN exhibited strong predictive ability in both the training and test sets, with AUC values of 0.907 (95% CI: 0.867\u0026ndash;0.946) and 0.873 (95% CI: 0.800-0.933), respectively. These AUC values were higher than those of the clinical model (Delong test, \u003cem\u003eP\u003c/em\u003e values all less than 0.001). DLRN was significantly superior to the clinical model, the single T1CE signature, and the single ADC signature in both the training and test sets (Supplementary Table S6). Further model comparison indicated that the constructed nomogram was superior to the manual radiomics model in all datasets, significantly superior to the deep learning feature model in the training set.\u003c/p\u003e \u003cp\u003eIn addition, the calibration curve of the diagnostic model showed good consistency between the observed results and the predicted results in both datasets (Hosmer-Lemeshow test, \u003cem\u003eP\u003c/em\u003e values were 0.966 and 0.713, respectively) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB, decision curve analysis revealed that the DLRN model achieves superior net clinical benefit across a wide range of thresholds compared to other models when intervening only on predicted high-risk patients.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe DLRN model achieved significantly higher accuracy (85.6%) and AUC (0.873) on the test set compared to the two assessors (accuracy: 63.3% and 68.9%; AUC: 0.660 and 0.686) (Supplementary Table S2). Specifically, the DLRN model indicated the greatest advantage in terms of sensitivity, indicating that it can more effectively identify TP cases and reduce false negative rates (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05). In the test set, the DLRN model significantly improved the diagnostic performance of radiologists C (Rc) and D (Rd), with net reclassification improvement (NRI) values of 1.102 (95% CI: 0.685\u0026ndash;1.490) and 1.298 (95% CI: 0.943\u0026ndash;1.625), respectively (Supplementary Table S7).\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eThis study presents a DLRN model that integrates ConvNeXt Tiny\u0026ndash;based deep learning features, radiomics, and clinical indicators to distinguish TP from PsP using conventional MRI. By capturing both imaging patterns and patient molecular background, the model enables accurate early risk stratification, helping to avoid delayed or excessive treatment and offering greater clinical value than visual assessment alone.\u003c/p\u003e \u003cp\u003eThe manual radiomics model in this study showed limited performance in differentiating TP from PsP (test AUC\u0026thinsp;=\u0026thinsp;0.617). Similarly, a T1CE-based texture model achieved only moderate accuracy (ACC: 72.78%)(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). Integrating clinical features has been shown to improve performance (AUC\u0026thinsp;=\u0026thinsp;0.729), suggesting that imaging features alone are insufficient(\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). Although radiomics from advanced functional MRI sequences (e.g., DWI, PWI) and multiparametric strategies can provide additional value (AUC up to 0.85)(\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e), the pure manual radiomics model in this study still exhibited limited predictive efficacy in the test set(\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe deep learning feature model constructed in this study demonstrated excellent discriminatory efficacy, with an AUC of 0.814 in the test set. In recent years, CNN, ViT, and their variants have been widely applied in differentiating glioma progression. For instance, Bacchi et al. achieved preliminary validation based on a multi-sequence CNN model in a small sample (accuracy 0.82), suggesting the potential of multi-sequence fusion(\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). Another study used a ViT model on an open dataset to confirm the baseline efficacy of a pure deep learning model (AUC\u0026thinsp;\u0026asymp;\u0026thinsp;0.717)(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Additionally, a study attempted to use complex architectures, such as CNN-LSTM, to process multiple sequences and achieved degrees of improvement in small samples (AUC 0.81)(\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). Compared to the above studies, this study has systematically optimized the following aspects: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Data level: A large multicenter cohort was adopted, thereby enhancing the model\u0026rsquo;s generalization reliability; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) Model architecture: The ConvNeXt Tiny network with superior performance was employed, overcoming the limited representational capacity of traditional CNNs(\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e); (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Strategy level: Deep learning features, traditional radiomics features, and key clinical molecular markers were innovatively integrated, achieving performance improvement based on comprehensive information.\u003c/p\u003e \u003cp\u003eThis study developed and validated a nomogram model integrating deep learning imaging features, conventional imaging biomarkers, and key clinical molecular indicators, which demonstrated excellent performance in distinguishing true progression from pseudoprogression in gliomas. A recent study indirectly confirms the universal validity and strong potential of the multimodal fusion strategy in neuro-tumor image analysis(\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). Using the nomogram proposed in this study, clinicians can directly convert the clinical features of patients, multi-sequence imaging biomarker signatures, and other indicators into specific scores. The individual score and the predicted probability of TP can be obtained by simply adding the scores of all indicators. Through rigorous auxiliary diagnosis experiments, this study proved that with the assistance of the DLRN model, the diagnostic accuracy of both radiologists increased to 80%, and the sensitivity reached 83.1% and 76.9%, respectively. This finding confirms that the DLRN model not only serves as an independent diagnostic tool but also provides an effective clinical decision support system.\u003c/p\u003e \u003cp\u003eThis study has several limitations. As a retrospective multicenter analysis, variations in scanning parameters and potential selection bias may have influenced performance despite standardized processing. The limited number of centers may restrict robustness, warranting future validation across more institutions, scanners, and protocols. Moreover, the cohort mainly included diagnostically challenging cases biased toward TP, which may limit generalizability to all patients with suspected progression.\u003c/p\u003e"},{"header":"5 Conclusions","content":"\u003cp\u003eThis study developed a ConvNeXt Tiny\u0026ndash;based multimodal DLRN model that outperformed single-modal models in distinguishing TP from PsP in gliomas and significantly improved radiologists\u0026rsquo; diagnostic accuracy with model assistance.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003e7.1 Ethics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective study was conducted in accordance with the ethical standards of the Declaration of Helsinki and was approved by the Ethics Review Board of the XXX Hospital of XXX (Ethics Number: XXXXXX). The requirement for written informed consent from participants was waived by the approving ethics committee due to the retrospective nature of the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.2 Consent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe need for written informed consent was waived with the confirmation of patient data confidentiality by the institutional Ethics Committee for this retrospective study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.3 Availability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are not publicly available due to privacy or ethical restrictions, but are available on request from the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.4 Competing interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.5 Funding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by Qinhuangdao S\u0026amp;T Plan Program (grant number: 202301A212)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.6 Author Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eT.Z.: Conceptualization, Writing \u0026ndash; original draft. L.Y.: Data curation, Writing \u0026ndash; original draft. D.Z.: Methodology. J.D.: Funding acquisition. X.L.: Formal analysis. S.W.: Visualization. X.W.: Supervision. Q.S.: Software. D.L.: Writing - Reviewing and Editing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.7 Acknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.8 Clinical trial number\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWeller M, Wen PY, Chang SM, Dirven L, Lim M, Monje M, Reifenberger G, Glioma (2024) Nat reviews Disease primers 10(1):33. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41572-024-00516-y\u003c/span\u003e\u003cspan address=\"10.1038/s41572-024-00516-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKowalczyk A, Zarychta J, Marszołek A, Zawitkowska J, Lejman M (2024) Chimeric Antigen Receptor T Cell and Chimeric Antigen Receptor NK Cell Therapy in Pediatric and Adult High-Grade Glioma-Recent Advances. Cancers 16(3). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/cancers16030623\u003c/span\u003e\u003cspan address=\"10.3390/cancers16030623\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWen PY, van den Bent M, Youssef G, Cloughesy TF, Ellingson BM, Weller M, Galanis E, Barboriak DP, de Groot J, Gilbert MR, Huang R, Lassman AB, Mehta M, Molinaro AM, Preusser M, Rahman R, Shankar LK, Stupp R, Villanueva-Meyer JE, Wick W, Macdonald DR, Reardon DA, Vogelbaum MA, Chang SM (2023) RANO 2.0: Update to the Response Assessment in Neuro-Oncology Criteria for High- and Low-Grade Gliomas in Adults. J Clin oncology: official J Am Soc Clin Oncol 41(33):5187\u0026ndash;5199. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1200/jco.23.01059\u003c/span\u003e\u003cspan address=\"10.1200/jco.23.01059\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHygino da Cruz LC Jr., Rodriguez I, Domingues RC, Gasparetto EL, Sorensen AG (2011) Pseudoprogression and pseudoresponse: imaging challenges in the assessment of posttreatment glioma. AJNR Am J Neuroradiol 32(11):1978\u0026ndash;1985. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3174/ajnr.A2397\u003c/span\u003e\u003cspan address=\"10.3174/ajnr.A2397\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdalla G, Hammam A, Anjari M, D'Arco DF, Bisdas DS (2020) Glioma surveillance imaging: current strategies, shortcomings, challenges and outlook. BJR open 2(1):20200009. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1259/bjro.20200009\u003c/span\u003e\u003cspan address=\"10.1259/bjro.20200009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCalmon R, Puget S, Varlet P, Dangouloff-Ros V, Blauwblomme T, Beccaria K, Grevent D, Sainte-Rose C, Castel D, Debily MA, Dufour C, Bolle S, Dhermain F, Saitovitch A, Zilbovicius M, Brunelle F, Grill J, Boddaert N (2018) Cerebral blood flow changes after radiation therapy identifies pseudoprogression in diffuse intrinsic pontine gliomas. Neurooncology 20(7):994\u0026ndash;1002. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/neuonc/nox227\u003c/span\u003e\u003cspan address=\"10.1093/neuonc/nox227\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBogsrud TV, Londalen A, Brandal P, Leske H, Panagopoulos I, Borghammer P, Bach-Gansmo T (2019) 18F-Fluciclovine PET/CT in Suspected Residual or Recurrent High-Grade Glioma. Clin Nucl Med 44(8):605\u0026ndash;611. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/rlu.0000000000002641\u003c/span\u003e\u003cspan address=\"10.1097/rlu.0000000000002641\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReddy S, Lung T, Muniyappa S, Hadley C, Templeton B, Fritz J, Boulter D, Shah K, Singh R, Zhu S, Matsui JK, Palmer JD (2025) Radiomics and Radiogenomics in Differentiating Progression, Pseudoprogression, and Radiation Necrosis in Gliomas. Biomedicines 13(7). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/biomedicines13071778\u003c/span\u003e\u003cspan address=\"10.3390/biomedicines13071778\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, Xu Q, Fan X, Cheng X, Tao J, Lu H, Lin Q, Zhang J, Qian J (2025) Multi-Sequence MRI radiomics model for discrimination of recurrence and pseudoprogression in gliomas. Eur J Radiol 194:112508. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ejrad.2025.112508\u003c/span\u003e\u003cspan address=\"10.1016/j.ejrad.2025.112508\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMayerhoefer ME, Materka A, Langs G, H\u0026auml;ggstr\u0026ouml;m I, Szczypiński P, Gibbs P, Cook G (2020) Introduction to Radiomics. Journal of nuclear medicine: official publication. Soc Nuclear Med 61(4):488\u0026ndash;495. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2967/jnumed.118.222893\u003c/span\u003e\u003cspan address=\"10.2967/jnumed.118.222893\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTripathi PC, Bag S (2023) An Attention-Guided CNN Framework for Segmentation and Grading of Glioma Using 3D MRI Scans. IEEE/ACM transactions on computational biology and bioinformatics. ;20(3):1890\u0026ndash;1904. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/tcbb.2022.3220902\u003c/span\u003e\u003cspan address=\"10.1109/tcbb.2022.3220902\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheng J, Liu J, Kuang H, Wang JA, Fully Automated (2022) Multimodal MRI-Based Multi-Task Learning for Glioma Segmentation and IDH Genotyping. IEEE Trans Med Imaging 41(6):1520\u0026ndash;1532. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/tmi.2022.3142321\u003c/span\u003e\u003cspan address=\"10.1109/tmi.2022.3142321\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu C, Peng Y, Zhu W, Chen Z, Li J, Tan W, Zhang Z, Chen X (2022) An automated approach for predicting glioma grade and survival of LGG patients using CNN and radiomics. Front Oncol 12:969907. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fonc.2022.969907\u003c/span\u003e\u003cspan address=\"10.3389/fonc.2022.969907\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWen PY, Macdonald DR, Reardon DA, Cloughesy TF, Sorensen AG, Galanis E, Degroot J, Wick W, Gilbert MR, Lassman AB, Tsien C, Mikkelsen T, Wong ET, Chamberlain MC, Stupp R, Lamborn KR, Vogelbaum MA, van den Bent MJ, Chang SM (2010) Updated response assessment criteria for high-grade gliomas: response assessment in neuro-oncology working group. J Clin oncology: official J Am Soc Clin Oncol 28(11):1963\u0026ndash;1972. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1200/jco.2009.26.3541\u003c/span\u003e\u003cspan address=\"10.1200/jco.2009.26.3541\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Sutter S, Wuts J, Geens W, Vanbinst AM, Duerinck J, Vandemeulebroucke J (2024) Modality redundancy for MRI-based glioblastoma segmentation. Int J Comput Assist Radiol Surg 19(10):2101\u0026ndash;2109. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s11548-024-03238-4\u003c/span\u003e\u003cspan address=\"10.1007/s11548-024-03238-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNigam S, Dheeraj A, Sachan H, Marwaha S (2025) Automated weed classification using attention-embedded ConvNeXtV2 architecture. Procedia Comput Sci 260:291\u0026ndash;299. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.procs.2025.03.204\u003c/span\u003e\u003cspan address=\"10.1016/j.procs.2025.03.204\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Mao H, Wu C-Y, Feichtenhofer C, Darrell T, Xie S A ConvNet for the 2020s. arXiv 2022. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/abs/2201.03545\u003c/span\u003e\u003cspan address=\"https://arxiv.org/abs/2201.03545\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun YZ, Yan LF, Han Y, Nan HY, Xiao G, Tian Q, Pu WH, Li ZY, Wei XC, Wang W, Cui GB (2021) Differentiation of Pseudoprogression from True Progressionin Glioblastoma Patients after Standard Treatment: A Machine Learning Strategy Combinedwith Radiomics Features from T(1)-weighted Contrast-enhanced Imaging. BMC Med Imaging 21(1):17. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12880-020-00545-5\u003c/span\u003e\u003cspan address=\"10.1186/s12880-020-00545-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAri AP, Akkurt BH, Musigmann M, Mammadov O, Bl\u0026ouml;mer DA, Kasap DNG, Henssen D, Nacul NG, Sartoretti E, Sartoretti T, Backhaus P, Thomas C, Stummer W, Heindel W, Mannil M (2022) Pseudoprogression prediction in high grade primary CNS tumors by use of radiomics. Sci Rep 12(1):5915. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-022-09945-9\u003c/span\u003e\u003cspan address=\"10.1038/s41598-022-09945-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim JY, Park JE, Jo Y, Shim WH, Nam SJ, Kim JH, Yoo RE, Choi SH, Kim HS (2019) Incorporating diffusion- and perfusion-weighted MRI into a radiomics model improves diagnostic performance for pseudoprogression in glioblastoma patients. Neurooncology 21(3):404\u0026ndash;414. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/neuonc/noy133\u003c/span\u003e\u003cspan address=\"10.1093/neuonc/noy133\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSala E, Mema E, Himoto Y, Veeraraghavan H, Brenton JD, Snyder A, Weigelt B, Vargas HA (2017) Unravelling tumour heterogeneity using next-generation imaging: radiomics, radiogenomics, and habitat imaging. Clin Radiol 72(1):3\u0026ndash;10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.crad.2016.09.013\u003c/span\u003e\u003cspan address=\"10.1016/j.crad.2016.09.013\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBacchi S, Zerner T, Dongas J, Asahina AT, Abou-Hamden A, Otto S, Oakden-Rayner L, Patel S (2019) Deep learning in the detection of high-grade glioma recurrence using multiple MRI sequences: A pilot study. J Clin neuroscience: official J Neurosurgical Soc Australasia 70:11\u0026ndash;13. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jocn.2019.10.003\u003c/span\u003e\u003cspan address=\"10.1016/j.jocn.2019.10.003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGomaa A, Huang Y, Stephan P, Breininger K, Frey B, D\u0026ouml;rfler A, Schnell O, Delev D, Coras R, Donaubauer AJ, Schmitter C, Stritzelberger J, Semrau S, Maier A, Bayer S, Sch\u0026ouml;necker S, Heiland DH, Hau P, Gaipl US, Bert C, Fietkau R, Schmidt MA, Putz F (2025) A self-supervised multimodal deep learning approach to differentiate post-radiotherapy progression from pseudoprogression in glioblastoma. Sci Rep 15(1):17133. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-02026-7\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-02026-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee J, Wang N, Turk S, Mohammed S, Lobo R, Kim J, Liao E, Camelo-Piragua S, Kim M, Junck L, Bapuraj J, Srinivasan A, Rao A (2020) Discriminating pseudoprogression and true progression in diffuse infiltrating glioma using multi-parametric MRI data through deep learning. Sci Rep 10(1):20331. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-020-77389-0\u003c/span\u003e\u003cspan address=\"10.1038/s41598-020-77389-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBijari S, Rezaeijo SM, Sayfollahi S, Rahimnezhad A, Heydarheydari S (2025) Development and validation of a robust MRI-based nomogram incorporating radiomics and deep features for preoperative glioma grading: a multi-center study. Quant imaging Med Surg 15(2):1125\u0026ndash;1138. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21037/qims-24-1543\u003c/span\u003e\u003cspan address=\"10.21037/qims-24-1543\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":true,"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":"glioma, deep learning, pseudo-progression, true progression","lastPublishedDoi":"10.21203/rs.3.rs-8643558/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8643558/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e \u003cp\u003eDifferentiating true progression (TP) from pseudoprogression (PsP) in glioma is challenging due to overlapping enhancement patterns on conventional MRI. Therefore, a reliable noninvasive approach integrating imaging heterogeneity is needed to improve TP/PsP discrimination.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis multicenter retrospective study included 294 patients with true progression (TP, n\u0026thinsp;=\u0026thinsp;208) or pseudoprogression (PsP, n\u0026thinsp;=\u0026thinsp;86). Baseline multiparametric MRI was analyzed. Traditional radiomics and deep learning features extracted using a pre-trained ConvNeXt Tiny network were selected through reproducibility, redundancy, and LASSO analyses to construct imaging signatures, which were combined with clinical factors to develop a deep learning radiomics nomogram (DLRN). Model performance was evaluated using ROC analysis, calibration curves, and decision curve analysis, and compared with radiologists\u0026rsquo; assessments.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe DLRN demonstrated excellent predictive efficacy, achieving an area under the curve (AUC) of 0.873 in the test set. Its performance significantly surpassed that of any individual signature (DeLong test, P\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and the independent assessments of two senior radiologists. The model exhibited good calibration, and decision curve analysis confirmed its superior clinical net benefit across a wide range of threshold probabilities. When used as a decision-support tool, the nomogram significantly and consistently improved both radiologists' diagnostic performance, yielding a net reclassification improvement greater than 1.1 in both the training and test sets (all P\u0026thinsp;\u0026lt;\u0026thinsp;0.01).\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThe deep learning imaging biomarker nomogram demonstrated excellent performance in differentiating TP from PsP in gliomas, outperforming traditional methods and radiologists, and effectively assisting clinical decision-making.\u003c/p\u003e","manuscriptTitle":"Integrating ConvNeXt Tiny and Radiomics in a Nomogram to Differentiate True Progression from Pseudoprogression in Glioma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-28 03:09:37","doi":"10.21203/rs.3.rs-8643558/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":"a9273368-bb0b-4142-b09d-ea442b7ada81","owner":[],"postedDate":"January 28th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-04T12:40:06+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-28 03:09:37","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8643558","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8643558","identity":"rs-8643558","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}