Abstract
Background
To explore the relationship between relaxed T2-FLAIR mismatch (RT2FM) sign, fractal dimension (FD) of tumor contour and IDH mutation status, and construct models for IDH mutation prediction in non-enhancing gliomas.
Methods
This retrospective study enrolled 364 patients with non-enhancing gliomas from two independent cohorts: cohort A (n = 267) for training & internal validation set and cohort B (n = 97) for external testing set. RT2FM, FD, and other MRI semantic features were extracted. Boruta and least absolute shrinkage and selection operator algorithms were employed to select intersecting features. Four machine learning models were constructed using the intersecting features and their diagnostic performance was evaluated.
Results
The RT2FM sign predicted IDH mutation with an accuracy, sensitivity, and specificity of 0.969, 0.650, and 0.921 in cohort A, and 0.924, 0.419, and 0.762 in cohort B, respectively. In cohort A, FD were significantly higher in IDH wild-type than in IDH mutant groups (1.264 vs. 1.190; P < 0.001). Using an FD cutoff value of 1.225, the area under the curve (AUC) and accuracy for predicting IDH mutation were 0.884 and 0.839, respectively. Among models constructed using four intersecting features (FD, RT2FM, multifocal/multicentric, and tumor location), XGBoost demonstrated the optimal predictive performance, with AUCs of 0.974, 0.968 and 0.895 in the training, internal validation and external testing set, respectively.
Conclusions
RT2FM and FD can provide informative imaging biomarkers for predicting IDH mutation. The XGBoost model constructed by these features demonstrated favorable diagnostic performance for IDH mutation prediction in non-enhancing gliomas.
Clinical trial number
Not applicable.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12880-025-02140-y.
Keywords: Isocitrate dehydrogenase, T2-FLAIR mismatch, Magnetic resonance imaging, Gliomas, Fractal analysis
Introduction
Diffuse glioma is the most common primary malignant brain tumor in adults. The 2021 World Health Organization (WHO) classification of tumors of central nervous system integrates molecular profiling with histopathology for the precise classification of diffuse glioma [1]. Isocitrate dehydrogenase (IDH) mutation status is established as a pivotal molecular biomarker for therapeutic decision-making and prognostic stratification [2].
Magnetic resonance imaging (MRI) is an important tool for gliomas diagnosis, with contrast enhancement on contrast-enhanced T1-weighted imaging (T1CE) serving as a radiographic biomarker indicative of aggressive biological behavior [3, 4]. However, subsequent evidence has confirmed that non-enhancing gliomas are a large and heterogeneous group. Of these, approximately 18% are IDH wild-type (IDHwt) [5], and exhibiting aggressive biological behavior. Therefore, preoperative prediction of IDH mutation is important in non-enhancing adult-type gliomas.
Glioma IDH mutation status drives tumor cell proliferative heterogeneity and spatial architecture [6], radiologically manifesting as MRI signal heterogeneity and irregular contours. Consequently, accurate quantification of these MRI-derived heterogeneity and morphological irregularity is essential for predicting IDH mutation status.
Functional MRI enables IDH mutation prediction through quantification of intratumoral heterogeneity including cellular density, perfusion, and metabolite products [7–9]. Its clinical implementation is constrained by prolonged acquisition times, economic burdens, and inter-institutional heterogeneity in acquisition parameters and post-processing protocols. In contrast, the T2-fluid-attenuated inversion recovery mismatch (T2FM) sign is a promising radiogenomic biomarker detected on routine MRI with 100% specificity for predicting IDH mutant (IDHmut) astrocytomas [10]. However, its stringent diagnostic criteria result in limited predictive sensitivity and suboptimal interobserver agreement, significantly constraining clinical applicability [11].
Traditional morphological parameters (e.g., tumor volume, maximal diameter) based on Euclidean geometry provide only coarse approximations of contour irregularity. While radiomic features like sphericity, compactness, and sphericity enable quantitative assessment, their limited intuitive biological interpretability and requirement for complex computational pipelines hinder clinical adoption [12]. In contrast, the tumor-brain interface’s inherent fractal properties offer an anatomically grounded framework [13], where fractal dimension (FD) emerging as a computationally metric that effectively quantifies contour irregularity without sophisticated mathematical operations [14].
In summary, previous studies on glioma intratumoral heterogeneity offered limited clinical utility and lacked effective methods for assessing tumor margin irregularity. Consequently, developing an IDH mutation prediction framework that integrates internal heterogeneity and contour irregularity indices, while ensuring clinical feasibility and interpretability, is imperative. In our study, we optimized T2FM to enhance its sensitivity, quantified contour irregularity using FD, and constructed an IDH prediction model combining FD with T2FM, accounting for both “internal heterogeneity” and “marginal invasiveness”.
Materials and methods
Patients
This study was approved by the institutional review board of Tangdu Hospital (IRB No. HG-202507-05) and Xi’an People’s Hospital (IRB No. 20230904). All procedures performed in this study involving human participants were in accordance with the ethical standards of the Declaration of Helsinki. Due to the retrospective nature of this study, the requirement for informed consent was waived.
Potentially eligible patients with pathologically confirmed gliomas were consecutively enrolled according to the inclusion and exclusion criteria. The inclusion criteria were: (i) age ≥ 18 years; (ii) preoperative MRI scan was performed; (iii) clinicopathologic information was well documented, including age, gender, WHO grade and IDH mutation status. The exclusion criteria were: (i) patient underwent treatment before MRI; (ii) discernible enhancement in lesion was observed on T1CE; (iii) MRI image quality was unsatisfactory due to susceptibility or motion artifacts; (iv) missing in any of the four routine sequences, including T1 weighted imaging (T1WI), T2 weighted imaging (T2WI), FLAIR and T1CE.
Finally, 267 patients from the Tangdu Hospital between March 2017 and May 2023 were included in the training and internal validation set, and 97 patients from the Xi’an People’s Hospital between February 2018 and May 2022 were used as the external testing set. The patient selection process is depicted in Fig. 1.
Fig. 1.
Patient selection flowchart
Histopathologic analysis
Histopathological and molecular analyses were in accordance with the 2021 WHO classification of tumors of the central nervous System. Sanger sequencing and next-generation sequencing were used to determine IDH mutation status. The lp/19q status was analyzed using fluorescent in situ hybridization or next-generation sequencing.
MRI acquisitions
MRI studies were performed at two institutions using either a 3.0 T or a 1.5 T unit from different scanners, with various parameters to reflect real inter-center heterogeneity. The brain tumor imaging protocol included T1WI, T2WI, FLAIR, and T1CE. The detailed acquisition parameters are provided in Supplementary Table S1.
Image analysis
Relaxed T2FM (RT2FM) definition
Diagnostic criteria for the relaxed T2FM sign: A lesion is classified as T2FM-positive if it displays focal region of T2WI hyperintensity with corresponding FLAIR hypointensity. Crucially, two points merit specification: First, the extent of mismatch region does not require involvement of the entire tumor region, and heterogeneous T2WI within this area is permissible. Second, cystic gliomas are classified as T2FM-positive. Representative cases are shown in Fig. S1.
MRI semantic feature assessment
Prior to feature assessment, all images were anonymized and radiologists were blinded to clinical history, pathological diagnosis, and molecular genotype of tumor. Two radiologists (Z.Z.L. and L.S.H., with 5 and 8 years of brain tumor diagnostic experience, respectively) independently evaluated MRI semantic features including RT2FM, tumor location, multicentric/multifocal, hyperintensity on T1WI, T2WI homogeneity, deep white matter invasion, cyst, cortical involvement, tumor borders, and sinuous, wave-like intratumoral-wall (SWITW) [15]. In cases of inter-reader discrepancy, consensus was achieved through consultation with a third board-certified radiologist (X.Y.B., with 20 years’ neuroimaging experience). Following a 3-month washout period, 30 randomly selected cases were reassessed for RT2FM sign by a senior radiologist. Intra- and inter-observer agreements of RT2FM were subsequently calculated.
Tumor segmentation and fractal analysis
Two radiologists (X.G. and S.J., with 4 and 9 years of brain tumor diagnostic experience respectively) manually segmented tumor regions of interest (ROI) on axial FLAIR images, blinded to IDH mutation status. Previous studies demonstrated that tumor segmentation using fewer slices can effectively capture molecular subtype-associated heterogeneity in gliomas while reducing segmentation time and enhancing clinical feasibility [16, 17]. In this study, ROIs were delineated on three adjacent consecutive slices centered on the slice demonstrating the maximal tumor diameter using open-source software ITK-SNAP (version 4.0.1; http://itk-snap.org).
The 2D FD values were calculated from the segmented masks using box-counting algorithms via Python [18]. As the optimal box size was unknown, a number of different 2D box sizes, ranging from 20 to 27 isotropic pixels, were adopted to compute the 2D FD [19–21]. The mean FD across the three slices was calculated per patient and utilized for subsequent analysis. Details of FD calculations are available in Supplementary S1.
Model development and validation for IDH mutation prediction
Patients from Tangdu Hospital were randomly allocated to the training and internal validation sets at a 7:3 ratio for model construction and evaluation, while cases from Xi’an People’s Hospital served as an independent external testing set. In the training set, clinical and MRI semantic features underwent dual feature selection via least absolute shrinkage and selection operator (LASSO) regression and Boruta algorithm, with intersecting features retained as IDH mutation-associated predictors [22]. Subsequently, intersecting features were used for machine learning models construction, including light gradient boosting machine (LightGBM), logistic regression (LR), extreme gradient boosting (XGBoost) and support vector machine (SVM). Random search and manual fine-tuning with 5-fold cross-validation were used to determine the optimal hyperparameters for each model. Finally, SHapley Additive exPlanations (SHAP) was used to assess the significance of each feature in the model [23].
Statistical analysis
Normality of the variables was assessed using the Shapiro-Wilk test. Quantitative variables were expressed as mean ± standard deviation for normal distribution or median with interquartile range (IQR) for non-normal distribution. Categorical variables were expressed as numbers and percentages. The Student’s t-test or one-way ANOVA test was used for continuous variables, and the Mann–Whitney U-test or Kruskal–Wallis H test was applied for nonparametric data. For comparative analyses of categorical variables in two groups, the chi-square test or Fisher’s exact test was used. Cohen’s Kappa analysis was used to assess the intra- and inter-observer agreement of RT2FM. Inter-class correlation coefficient (ICC) was calculated to evaluate the intra- and inter-observer agreement for FD. Receiver operating characteristic (ROC) curve was constructed to calculate area under the curve (AUC) and the optimal cut-off value was determined by the maximal Youden index. The performance of the IDH mutation prediction model was evaluated using multiple metrics, including AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy (ACC), F1 score, and Matthews correlation coefficient (MCC). The model’s fit was tested through calibration curve. The decision curve analysis (DCA) was utilized to assess the clinical utility of the model. Statistical analysis was performed using software (SPSS, version 26.0; MedCalc, version 15.0; R, version 4.3.2; Python, version 3.8.5). P < 0.05 was considered a statistically significant difference.
Results
Patient characteristics
Baseline characteristics is detailed in Table 1. In cohort A, 267 patients (median age, 44 years; IQR, 36–52 years; 157 males) were included, comprising 129 astrocytomas, 98 oligodendrogliomas and 40 glioblastomas. Cohort B included 97 patients (median age, 42 years, IQR, 33–51 years; 56 males) with 66 IDHmut gliomas (25 astrocytomas, 41 oligodendrogliomas) and 31 glioblastomas. IDHmut gliomas exhibited a younger age compared to IDHwt group (cohort A, P = 0.002; cohort B, P < 0.001).
Table 1.
Patient characteristics
| Characteristics | Cohort A: Training set & Internal validation set | Cohort B: External testing set | ||||
|---|---|---|---|---|---|---|
| IDHmut (n = 227) | IDHwt (n = 40) | P | IDHmut (n = 66) | IDHwt (n = 31) | P | |
| Age(y) § | 44 (36, 51) | 52 (39, 61) | 0.002‡ | 39 (31,47) | 51 (42, 62) | < 0.001‡ |
| Gender | 0.380Ψ | 0.403Ψ | ||||
| Male | 136 (59.91) | 21 (52.50) | 40 (60.61) | 16 (51.61) | ||
| Female | 91 (40.09) | 19 (47.50) | 26 (39.39) | 15 (48.39) | ||
| Histopathology | NA | NA | ||||
| Glioblastoma | 0 (0.00) | 40 (100.00) | 0 (0.00) | 31 (100.00) | ||
| Astrocytoma | 129 (56.83) | 0 (0.00) | 25 (37.88) | 0 (0.00) | ||
| Oligodendroglioma | 98 (43.17) | 0 (0.00) | 41 (62.12) | 0 (0.00) | ||
| WHO Grade | < 0.001Ψ | < 0.001Ψ | ||||
| Grade 2 | 164 (72.25) | 0 (0.00) | 42 (63.64) | 0 (0.00) | ||
| Grade 3 | 54 (23.79) | 0 (0.00) | 22 (33.33) | 0 (0.00) | ||
| Grade 4 | 9 (3.96) | 40 (100.00) | 2 (3.03) | 31 (100.00) | ||
Note. — Unless otherwise indicated, data are numbers of patients, and data in parentheses are percentages. IDH = isocitrate dehydrogenase, IDHmut = IDH mutant, IDHwt = IDH wild type, NA = not available
§Data are the median, and data in parentheses are the interquartile range
‡Mann-Whitney U test
ΨPearson’s Chi-squared test
MRI semantic features
Table 2 presents MRI semantic feature differences between IDHmut and IDHwt groups across cohorts A and B. In cohort A, IDHmut gliomas exhibited frontal lobe predilection, well-defined tumor border, cortical involvement, heterogeneous T2WI, hyperintensity on T1WI and a positive SWITW sign. In contrast, IDHwt gliomas tended to show blurred borders, multifocal/multicentric and deep white matter invasion. Similar trends were observed in cohort B. No statistically differences were observed in cyst between IDHmut and IDHwt groups across both cohorts (all P > 0.05). All semantic features demonstrated substantial to almost perfect inter- and intra-observer agreement, as detailed in Supplementary Table S2.
Table 2.
MRI semantic feature comparisons between idhmut and IDHwt groups
| Characteristics | Cohort A: Training set & Internal validation set | Cohort B: External testing set | ||||
|---|---|---|---|---|---|---|
| IDHmut (n = 227) | IDHwt (n = 40) | P | IDHmut (n = 66) | IDHwt (n = 31) | P | |
| Location | < 0.001ξ | < 0.001ξ | ||||
| Frontal | 157 (69.16) | 10 (25.00) | 42 (63.64) | 9 (29.03) | ||
| Temporal-insular | 56 (24.67) | 6 (15.00) | 19 (28.79) | 6 (19.35) | ||
| Parieto-occipital | 12 (5.29) | 5 (12.50) | 4 (6.06) | 7 (22.58) | ||
| Other | 2 (0.88) | 19 (47.50) | 1 (1.51) | 9 (29.04) | ||
| RT2FM | < 0.001ξ | < 0.001Ψ | ||||
| Present | 220 (96.92) | 14 (35.00) | 61 (92.42) | 18 (58.06) | ||
| Absent | 7 (3.08) | 26 (65.00) | 5 (7.58) | 13 (41.94) | ||
| Multifocal/Multicentric | < 0.001ξ | < 0.001Ψ | ||||
| Present | 8 (3.52) | 22 (55.00) | 5 (7.58) | 11 (35.48) | ||
| Absent | 219 (96.48) | 18 (45.00) | 61 (92.42) | 20 (64.52) | ||
| Deep WM invasion | < 0.001ξ | 0.001Ψ | ||||
| Present | 10 (4.41) | 14 (35.00) | 7 (10.61) | 12 (38.71) | ||
| Absent | 217 (95.59) | 26 (65.00) | 59 (89.39) | 19 (61.29) | ||
| Cyst (s) | 0.642Ψ | 0.708ξ | ||||
| Present | 47 (20.70) | 7 (17.50) | 5 (7.58) | 3 (9.68) | ||
| Absent | 180 (79.30) | 33 (82.50) | 61 (92.42) | 28 (90.32) | ||
| Cortical involvement | < 0.001ξ | 0.032ξ | ||||
| Present | 224 (98.68) | 31 (77.50) | 64 (96.97) | 26 (83.87) | ||
| Absent | 3 (1.32) | 9 (22.50) | 2 (3.03) | 5 (16.13) | ||
| T2WI homogeneity | < 0.001Ψ | 0.046Ψ | ||||
| Homogeneous | 23 (10.13) | 13 (32.50) | 13 (19.70) | 12 (38.71) | ||
| Heterogeneous | 204 (89.87) | 27 (67.50) | 53 (80.30) | 19 (61.29) | ||
| Tumor border | < 0.001Ψ | 0.001Ψ | ||||
| Well defined | 96 (42.29) | 4 (10.00) | 36 (54.55) | 6 (19.35) | ||
| Poorly defined | 131 (57.71) | 36 (90.00) | 30 (45.45) | 25 (80.65) | ||
| Hyperintensity on T1WI | 0.032Ψ | |||||
| Present | 65 (28.63) | 5 (12.50) | 21 (31.82) | 1 (3.23) | 0.002Ψ | |
| Absent | 162 (71.37) | 35 (87.50) | 45 (68.18) | 30 (96.77) | ||
| SWITW | 0.016Ψ | |||||
| Present | 56 (24.67) | 3 (7.50) | 22 (33.33) | 4 (12.90) | 0.034Ψ | |
| Absent | 171 (75.33) | 37 (92.50) | 44 (66.67) | 27 (87.10) | ||
Note. — Unless otherwise indicated, data are numbers of patients, and data in parentheses are percentages. IDH = isocitrate dehydrogenase, IDHmut = IDH mutant, IDHwt = IDH wild type, T1WI = T1 weighted imaging, T2WI = T2 weighted imaging, FLAIR = fluid-attenuated inversion recovery, RT2FM = relaxed T2-FLAIR mismatch, Deep WM invasion = deep white matter invasion, SWITW = sinuous wave-like intratumoral-wall
ξFisher’s exact test
ΨPearson’s Chi-squared test
RT2FM for IDH mutation prediction
Figure 2 depicts RT2FM distribution across histopathology subtypes (Sankey plot, left) and its diagnostic performance for predicting IDH mutation status (Radar chart, right). In cohort A, 234 cases showed positive RT2FM (astrocytoma, 127/129; oligodendroglioma, 93/98; glioblastoma, 14/40), while 33 cases were negative (astrocytoma, 2/129; oligodendroglioma, 5/98; glioblastoma, 26/40). In cohort B, positive RT2FM was observed in 79 cases (astrocytoma, 24/25; oligodendroglioma, 37/41; glioblastoma, 18/31), and 18 cases were negative (astrocytoma, 1/25; oligodendroglioma, 4/41; glioblastoma, 13/31). The presence, sensitivity, specificity, and accuracy of RT2FM for predicting IDH mutation were 0.876, 0.969, 0.650, and 0.921 in cohort A, and 0.814, 0.924, 0.419, and 0.762 in cohort B, respectively. The intra-observer and inter-observer agreement for RT2FM was almost perfect, with κ values of 0.867 and 0.824, respectively (Supplementary Table S2).
Fig. 2.
Sankey plots (left column) of RT2FM distribution across histopathology subtypes and radar plots (right column) of RT2FM performance for IDH mutation prediction in cohort A (A) and B (B). Abbreviations: IDH = isocitrate dehydrogenase, T2WI = T2 weighted image, FLAIR = fluid-attenuated inversion recovery, RT2FM = relaxed T2-FLAIR mismatch, PRE = presence, SEN = sensitivity, SPE = specificity, PPV = positive predictive value, NPV = negative predictive value, ACC = accuracy
FD for IDH mutation prediction
Representative cases of fractal analysis are shown in Fig. 3A. The FD measurements demonstrated excellent consistency, with intra-observer and inter-observer ICCs of 0.957 and 0.909, respectively (Fig. S2). In cohort A, IDHwt gliomas exhibited higher FD than the IDHmut group (IDHwt, 1.264 vs. IDHmut, 1.190; P < 0.001) (Fig. 3B). Further subgroup analysis revealed no significant difference in FD was observed between oligodendrogliomas (IDHmut-Noncodel) and astrocytomas (IDHmut-Codel) (Fig. 3C). When the cut-off value was set at 1.225, the AUC for predicting IDH mutation was 0.884, with a sensitivity of 0.837 and a specificity of 0.850 (Fig. 3D).
Fig. 3.
Fractal analysis. (A) Representative cases: images from left column to right column are as follows: axial T2WI, FLAIR and outline image of maximum cross-sectional tumor diameter. (B) Violin plot comparing FD between IDHwt and IDHmut groups. (C) Violin plot comparing FD for IDHmut-Noncodel, IDHmut-Codel and IDHwt groups. (D) The ROC curve of FD for predicting IDH mutation status. Abbreviations: IDH = isocitrate dehydrogenase, IDHmut = IDH mutant, IDHwt = IDH wild type (glioblastoma), IDHmut-Noncodel = IDH mutant and 1p/19q noncodeletion astrocytoma, IDHmut-Codel = IDH mutant and 1p/19q codeletion (oligodendroglioma), T2WI = T2 weighted imaging, FLAIR = fluid-attenuated inversion recovery, FD = fractal dimension, ROC = receiver operating characteristic, AUC = area under curve, NS = no significance, ****, P < 0.001
Model performance for IDH mutation prediction
Figure S3 illustrates the intersectional feature selection by LASSO and Boruta algorithm, identifying four critical features: FD, tumor location, multicentric/multifocal, and RT2FM. Utilizing these features, LightGBM, LR, XGBoost, and SVM were subsequently developed. As detailed in Table 3, XGBoost model exhibited superior performance, with AUCs of 0.974/0.968/0.895 and ACCs of 0.952/0.950/0.876 across training/internal validation/external testing sets. The XGBoost model demonstrated good calibration and net benefit in the training set, internal validation set, and external testing set (Fig. 4A). Figure 4B-C displays SHAP-based feature rankings in the XGBoost model: FD > RT2FM > multifocal/multicentric > location. Specifically, lower FD, positive RT2FM, and negative multifocal/multicentric status were significantly associated with IDH mutation.
Table 3.
Comparisons of model performance for IDH mutation prediction across training, internal validation and external testing set
| AUC (95%CI) | SEN | SPE | PPV | NPV | ACC | F1 score | MCC | |
|---|---|---|---|---|---|---|---|---|
| XGBoost | ||||||||
| Training set | 0.974 (0.936-1.000) | 0.956 | 0.929 | 0.987 | 0.788 | 0.952 | 0.971 | 0.825 |
| Internal validation set | 0.968 (0.925-1.000) | 0.941 | 0.833 | 0.970 | 0.714 | 0.950 | 0.955 | 0.728 |
| External testing set | 0.895 (0.820–0.970) | 0.955 | 0.710 | 0.875 | 0.880 | 0.876 | 0.913 | 0.708 |
| LightGBM | ||||||||
| Training set | 0.969 (0.941–0.996) | 0.893 | 0.893 | 0.979 | 0.595 | 0.893 | 0.934 | 0.672 |
| Internal validation set | 0.954 (0.901-1.000) | 0.912 | 0.833 | 0.969 | 0.625 | 0.900 | 0.939 | 0.665 |
| External testing set | 0.893 (0.826–0.959) | 0.803 | 0.774 | 0.883 | 0.649 | 0.794 | 0.841 | 0.599 |
| SVM | ||||||||
| Training set | 0.955 (0.909-1.000) | 0.994 | 0.821 | 0.969 | 0.958 | 0.968 | 0.981 | 0.870 |
| Internal validation set | 0.850 (0.702–0.998) | 0.985 | 0.583 | 0.931 | 0.875 | 0.925 | 0.957 | 0.677 |
| External testing set | 0.883 (0.810–0.957) | 0.984 | 0.580 | 0.833 | 0.947 | 0.856 | 0.903 | 0.664 |
| LR | ||||||||
| Training set | 0.946 (0.892-1.000) | 0.975 | 0.857 | 0.975 | 0.857 | 0.957 | 0.974 | 0.832 |
| Internal validation set | 0.968 (0.922-1.000) | 0.985 | 0.667 | 0.944 | 0.889 | 0.938 | 0.964 | 0.737 |
| External testing set | 0.876 (0.801–0.951) | 0.985 | 0.548 | 0.823 | 0.944 | 0.845 | 0.897 | 0.640 |
Note. — AUC = area under the receiver operating characteristic curve, CI = confidence interval, SEN = sensitivity, SPE = specificity, PPV = positive predictive value, NPV = negative predictive value, ACC = accuracy, MCC = matthews correlation coefficient, XGBoost = extreme gradient boosting, LightGBM = light gradient boosting machine, SVM = support vector machine, LR = logistic regression
Fig. 4.
Performance evaluation of XGBoost model. (A) Calibration curves and decision curve analysis of XGBoost model for IDH mutation prediction across the training, internal validation, and external testing sets. (B) Bee swarm plot of SHAP analysis for XGBoost model. (C) Feature importance ranking plot of the XGBoost model. Abbreviations: DCA = decision curve analysis, IDH = isocitrate dehydrogenase, T2WI = T2 weighted imaging, FLAIR = fluid-attenuated inversion recovery, RT2FM = relaxed T2-FLAIR mismatch, FD = fractal dimension, XGBoost = extreme gradient boosting, SHAP = SHapley Additive exPlanations
Discussion
Exploration of a noninvasive tool to predict IDH mutation status is important to guide clinical decision-making in non-enhancing adult-type diffuse gliomas. Radiologically, gliomas with distinct IDH mutation status exhibit differences in intratumoral signal and contour irregularity. In our study, we employed RT2FM and other MRI semantic features to classify intratumoral heterogeneity while utilizing FD to quantify contour irregularity. Ultimately, four models were constructed to predict IDH mutation status by integrating RT2FM, FD and MRI semantic features. XGBoost model exhibited optimal diagnostic performance, with AUCs of 0.974/0.968/0.895 across training/internal validation/external testing sets.
The classical T2FM requires a homogeneous T2WI signal and complete FLAIR signal attenuation (except for the high signal edge), resulting in reduced sensitivity for IDH mutation prediction in the real-world cohorts [24, 25]. In this study, we applied a relaxed T2FM criterion that neither confines mismatch region to the entire tumor area nor requires homogeneous T2WI signal. Notably, for tumors exhibiting prominent cystic gliomas, we adopted the criteria established by Lee et al. [26] to classify these cases as T2FM-positive, corresponding to IDHmut status. Following the aforementioned adjustments, our RT2FM demonstrated sensitivities of 0.969 and 0.924 in IDH mutation prediction across cohorts A and B, respectively. In alignment with our study design, Throckmorton et al. [11] demonstrated that incorporating T2WI heterogeneity (without restricting T2WI homogeneity) improved the sensitivity of T2FM for IDH prediction from 34% to 74%. Subsequent studies by Lasoki et al. [27] and Cho et al. [28] further quantified the spatial extent of mismatch involvement. By defining T2FM positive as a mismatch region involvement extent of ≥ 25% of tumor, these studies achieved sensitivities of 0.33–0.63 for IDH prediction while maintaining high specificity (0.92–1.00). However, the quantitative assessment of mismatch extent in these studies required complex postprocessing, and substantial interobserver variability, which may limit their clinical applicability. Although RT2FM improved the sensitivity for IDH mutation prediction, its specificity remained suboptimal, with values of 0.650 and 0.419 in cohorts A and B, respectively, likely attributable to local mismatch sign observed in glioblastoma [29]. Nevertheless, parameter complementarity in the multiparametric modeling framework mitigated this limitation. SHAP analysis further validated RT2FM as a critical predictive marker for IDH mutation prediction.
The FD offers both biological insight and clinical relevance in distinguishing between glioma subtypes. Biologically, FD quantitatively captures the irregularity of tumor contours, which directly mirrors underlying differences in cellular behavior and growth patterns [30]. Consistent with the findings of Saha et al. [31], our study also demonstrated a higher median FD value of GBMs compared to the IDHmut gliomas, indicating IDH-wild type glioblastomas greater spatial heterogeneity and more aggressive. This aligns with the known pathology of these tumors, where pronounced invasiveness and heterogeneous cell proliferation drive complex, irregular morphologies [32, 33]. Thus, FD serves as a non-invasive imaging biomarker that reflects the proliferation-invasion process of tumors. Clinically, the stratification of glioma subcomponents by FD has direct implications for prognosis. A lower FD in enhancing subcomponent correlates with their typically more circumscribed growth and slower progression, which are factors associated with a more favorable prognosis [34]. In summary, FD translates a visually complex morphological characteristic into an objective quantitative parameter that links tumor biology with clinical decision-making. Considering clinical applicability, our study employed manual segmentation limited to three consecutive slices centered on tumor’s largest cross-sectional area. This approach not only reduced radiologist’s time expenditure but also ensured reproducibility, thereby supporting potential clinical translation. Furthermore, SHAP-based feature rankings in the XGBoost model identified FD as the most critical variable for predicting IDH mutation in gliomas.
During model construction, we incorporated two MRI semantic features, tumor location and multicentric/multifocal presentation, alongside RT2FM and FD. Our findings indicated that IDHmut gliomas were predominantly located in the frontal lobe, whereas multicentric/multifocal lesions were more common in IDHwt cases, aligning with previous studies [35, 36]. Among the four constructed models for IDH mutation prediction, XGBoost achieved the highest performance, likely due to its gradient-boosted decision tree architecture, which efficiently handles nonlinear relationships, mitigates class imbalance, and ensures stable classification accuracy [37, 38]. This algorithm has been widely implemented in artificial intelligence applications and is increasingly utilized for tumor molecular subtyping and prognostic prediction [39, 40].
This study has several limitations. First, the retrospective design inherently carries potential selection bias. Second, despite incorporating data from two independent institutions, the sample size for non-enhancing GBM remained limited. Future multi-center studies with larger cohorts are necessary to validate the generalizability of the model. Third, tumor segmentation was conducted using only 3 consecutive imaging slices. Although this approach improved analytical efficiency, it may fail to fully capture the intratumoral heterogeneity of lesions with extensive infiltration and highly irregular morphologies. Fourth, the prediction of IDH mutation status was primarily based on conventional MRI, without comparative analysis incorporating functional MRI sequences. Finally, although the RT2FM demonstrated high sensitivity in predicting IDH mutation, its specificity was limited. Further refinement of the RT2FM criteria to improve specificity will be a focus of future research.
Conclusions
In our study, RT2FM significantly improved sensitivity for preoperative IDH mutation prediction in non-enhancing adult-type diffuse gliomas. Additionally, IDHwt gliomas exhibited higher FD compared to IDHmut counterparts. Furthermore, the XGBoost model integrating RT2FM, FD, and other MRI semantic features showed potential for IDH mutation prediction. However, validation for our findings in a prospective, multicenter cohort is warranted.
Supplementary information
Below is the link to the electronic supplementary material.
Acknowledgements
None.
Abbreviations
- AUC
area under the curve
- FD
fractal dimension
- FLAIR
fluid-attenuated inversion recovery
- IDH
isocitrate dehydrogenase
- IQR
interquartile range
- LASSO
least absolute shrinkage and selection operator
- MCC
matthews correlation coefficient
- SHAP
SHapley Additive exPlanations
- ROC
receiver operating characteristic
- T1CE
contrast-enhanced T1WI
- RT2FM
relaxed T2-FLAIR mismatch
- WHO
world health organization
Author contributions
Yu Han: Writing – review & editing, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Jin Zhang: Writing – review & editing, Writing – original draft, Methodology, Conceptualization. Yi-bin Xi: Writing – review & editing, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Si-jie Xiu: Writing – review & editing, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Yang Yang: Writing – review & editing, Supervision, Project administration, Methodology, Conceptualization. Yu-yao Wang: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization.
Funding
This study received financial support from the National Natural Science Foundation of China (No. 82102127 to YY and No. 82371936 to YBX).
Data availability
The datasets generated or analyzed during the study are not publicly available but are available from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
This study was approved by the institutional review board of Tangdu Hospital (IRB No. K-HG-202507-05) and Xi’an People’s Hospital (IRB No. 20230904). All procedures performed in this study involving human participants were in accordance with the ethical standards of the Declaration of Helsinki. Due to the retrospective nature of this study, the requirement for informed consent was waived.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Yu Han and Jin Zhang contributed equally to this work.
Contributor Information
Yang Yang, Email: yyang507@126.com.
Yu-yao Wang, Email: wyy2007154030@163.com.
References
- 1.Louis DN, Perry A, Wesseling P, Brat DJ, Cree IA, Figarella-Branger D, et al. The 2021 WHO classification of tumors of the central nervous system: a summary. Neuro Oncol. 2021;23(8):1231–51. 10.1093/neuonc/noab106. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Berger TR, Wen PY, Lang-Orsini M, Chukwueke UN. World health organization 2021 classification of central nervous system tumors and implications for therapy for Adult-Type gliomas: A review. JAMA Oncol. 2022;8(10):1493–501. 10.1001/jamaoncol.2022.2844. [DOI] [PubMed] [Google Scholar]
- 3.Brindle KM, Izquierdo-García JL, Lewis DY, Mair RJ, Wright AJ. Brain tumor imaging. J Clin Oncol. 2017;35(21):2432–8. 10.1200/jco.2017.72.7636. [DOI] [PubMed] [Google Scholar]
- 4.Garzón B, Emblem KE, Mouridsen K, Nedregaard B, Due-Tønnessen P, Nome T, et al. Multiparametric analysis of magnetic resonance images for glioma grading and patient survival time prediction. Acta Radiol. 2011;52(9):1052–60. 10.1258/ar.2011.100510. [DOI] [PubMed] [Google Scholar]
- 5.Kang KM, Song J, Choi Y, Park C, Park JE, Kim HS, et al. MRI scoring systems for predicting isocitrate dehydrogenase mutation and chromosome 1p/19q codeletion in Adult-type diffuse glioma lacking contrast enhancement. Radiology. 2024;311(2):e233120. 10.1148/radiol.233120. [DOI] [PubMed] [Google Scholar]
- 6.Moffet JJD, Fatunla OE, Freytag L, Kriel J, Jones JJ, Roberts-Thomson SJ, et al. Spatial architecture of high-grade glioma reveals tumor heterogeneity within distinct domains. Neurooncol Adv. 2023;5(1):vdad142. 10.1093/noajnl/vdad142. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Pruis IJ, Koene SR, van der Voort SR, Incekara F, Vincent A, van den Bent MJ, et al. Noninvasive differentiation of molecular subtypes of adult nonenhancing glioma using MRI perfusion and diffusion parameters. Neurooncol Adv. 2022;4(1):vdac023. 10.1093/noajnl/vdac023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Arzanforoosh F, van der Voort SR, Incekara F, Vincent A, Van den Bent M, Kros JM, et al. Microvasculature features derived from hybrid EPI MRI in non-enhancing adult-type diffuse glioma subtypes. Cancers (Basel). 2023;15(7). 10.3390/cancers15072135. [DOI] [PMC free article] [PubMed]
- 9.Wu M, Jiang T, Guo M, Duan Y, Zhuo Z, Weng J, et al. Amide proton transfer-weighted imaging and derived radiomics in the classification of adult-type diffuse gliomas. Eur Radiol. 2024;34(5):2986–96. 10.1007/s00330-023-10343-6. [DOI] [PubMed] [Google Scholar]
- 10.Patel SH, Poisson LM, Brat DJ, Zhou Y, Cooper L, Snuderl M, et al. T2-FLAIR Mismatch, an imaging biomarker for IDH and 1p/19q status in Lower-grade gliomas: A TCGA/TCIA project. Clin Cancer Res. 2017;23(20):6078–85. 10.1158/1078-0432.Ccr-17-0560. [DOI] [PubMed] [Google Scholar]
- 11.Throckmorton P, Graber JJ. T2-FLAIR mismatch in isocitrate dehydrogenase mutant astrocytomas: variability and evolution. Neurology. 2020;95(11):e1582–9. 10.1212/wnl.0000000000010324. [DOI] [PubMed] [Google Scholar]
- 12.Park JE, Kim D, Kim HS, Park SY, Kim JY, Cho SJ, et al. Quality of science and reporting of radiomics in oncologic studies: room for improvement according to radiomics quality score and TRIPOD statement. Eur Radiol. 2020;30(1):523–36. 10.1007/s00330-019-06360-z. [DOI] [PubMed] [Google Scholar]
- 13.Torres Hoyos F, Navarro RB, Vergara Villadiego J, Guerrero-Martelo M. Geometrical study of Astrocytomas through fractals and scaling analysis. Appl Radiat Isot. 2018;141:250–6. 10.1016/j.apradiso.2018.05.020. [DOI] [PubMed] [Google Scholar]
- 14.Lennon FE, Cianci GC, Cipriani NA, Hensing TA, Zhang HJ, Chen CT, et al. Lung cancer-a fractal viewpoint. Nat Rev Clin Oncol. 2015;12(11):664–75. 10.1038/nrclinonc.2015.108. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Li M, Wang J, Chen X, Dong G, Zhang W, Shen S, et al. The sinuous, wave-like intratumoral-wall sign is a sensitive and specific radiological biomarker for oligodendrogliomas. Eur Radiol. 2023;33(6):4440–52. 10.1007/s00330-022-09314-0. [DOI] [PubMed] [Google Scholar]
- 16.Kocak B, Durmaz ES, Ates E, Sel I, Turgut Gunes S, Kaya OK, et al. Radiogenomics of lower-grade gliomas: machine learning-based MRI texture analysis for predicting 1p/19q codeletion status. Eur Radiol. 2020;30(2):877–86. 10.1007/s00330-019-06492-2. [DOI] [PubMed] [Google Scholar]
- 17.Chu JP, Song YK, Tian YS, Qiu HS, Huang XH, Wang YL, et al. Diffusion kurtosis imaging in evaluating gliomas: different region of interest selection methods on time efficiency, measurement repeatability, and diagnostic ability. Eur Radiol. 2021;31(2):729–39. 10.1007/s00330-020-07204-x. [DOI] [PubMed] [Google Scholar]
- 18.Lopes R, Betrouni N. Fractal and multifractal analysis: a review. Med Image Anal. 2009;13(4):634–49. 10.1016/j.media.2009.05.003. [DOI] [PubMed] [Google Scholar]
- 19.Petrujkić K, Milošević N, Rajković N, Stanisavljević D, Gavrilović S, Dželebdžić D, et al. Computational quantitative MR image features - a potential useful tool in differentiating glioblastoma from solitary brain metastasis. Eur J Radiol. 2019;119:108634. 10.1016/j.ejrad.2019.08.003. [DOI] [PubMed] [Google Scholar]
- 20.Curtin L, Whitmire P, White H, Bond KM, Mrugala MM, Hu LS, Swanson KR. Shape matters: morphological metrics of glioblastoma imaging abnormalities as biomarkers of prognosis. Sci Rep. 2021;11(1):23202. 10.1038/s41598-021-02495-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Kim S, Park YW, Park SH, Ahn SS, Chang JH, Kim SH, Lee SK. Comparison of diagnostic performance of Two-Dimensional and Three-Dimensional fractal dimension and lacunarity analyses for predicting the meningioma grade. Brain Tumor Res Treat. 2020;8(1):36–42. 10.14791/btrt.2020.8.e3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Sun T, Liu J, Yuan H, Li X, Yan H. Construction of a risk prediction model for lung infection after chemotherapy in lung cancer patients based on the machine learning algorithm. Front Oncol. 2024;14:1403392. 10.3389/fonc.2024.1403392. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Muhammad D, Bendechache M. Unveiling the black box: A systematic review of explainable artificial intelligence in medical image analysis. Comput Struct Biotechnol J. 2024;24:542–60. 10.1016/j.csbj.2024.08.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Dagher SA, Lochner RH, Ozkara BB, Schomer DF, Wintermark M, Fuller GN, Ucisik FE. The T2-FLAIR mismatch sign in oncologic neuroradiology: History, current use, emerging data, and future directions. Neuroradiol J. 2024;37(4):441–53. 10.1177/19714009231212375. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Han Z, Chen Q, Zhang L, Mo X, You J, Chen L, et al. Radiogenomic association between the T2-FLAIR mismatch sign and IDH mutation status in adult patients with lower-grade gliomas: an updated systematic review and meta-analysis. Eur Radiol. 2022;32(8):5339–52. 10.1007/s00330-022-08607-8. [DOI] [PubMed] [Google Scholar]
- 26.Lee MD, Patel SH, Mohan S, Akbari H, Bakas S, Nasrallah MP, et al. Association of partial T2-FLAIR mismatch sign and isocitrate dehydrogenase mutation in WHO grade 4 gliomas: results from the respond consortium. Neuroradiology. 2023;65(9):1343–52. 10.1007/s00234-023-03196-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Lasocki A, Buckland ME, Drummond KJ, Wei H, Xie J, Christie M, et al. Conventional MRI features can predict the molecular subtype of adult grade 2–3 intracranial diffuse gliomas. Neuroradiology. 2022;64(12):2295–305. 10.1007/s00234-022-02975-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Cho NS, Sanvito F, Le VL, Oshima S, Teraishi A, Yao J, et al. Quantification of T2-FLAIR mismatch in nonenhancing diffuse gliomas using digital Subtraction. AJNR Am J Neuroradiol. 2024;45(2):188–97. 10.3174/ajnr.A8094. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Malik P, Soliman R, Chen YA, Munoz DG, Das S, Bharatha A, Mathur S. Patterns of T2-FLAIR discordance across a cohort of adult-type diffuse gliomas and deviations from the classic T2-FLAIR mismatch sign. Neuroradiology. 2024;66(4):521–30. 10.1007/s00234-024-03297-z. [DOI] [PubMed] [Google Scholar]
- 30.Sánchez J, Martín-Landrove M. Morphological and fractal properties of brain tumors. Front Physiol. 2022;13:878391. 10.3389/fphys.2022.878391. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Mazurowski MA, Clark K, Czarnek NM, Shamsesfandabadi P, Peters KB, Saha A. Radiogenomics of lower-grade glioma: algorithmically-assessed tumor shape is associated with tumor genomic subtypes and patient outcomes in a multi-institutional study with the cancer genome atlas data. J Neurooncol. 2017;133(1):27–35. 10.1007/s11060-017-2420-1. [DOI] [PubMed] [Google Scholar]
- 32.Wang J, Yi L, Kang QM, Zhou J, Chen TQ, Hugnot JP, Yu SC. Glioma invasion along white matter tracts: A dilemma for neurosurgeons. Cancer Lett. 2022;526:103–11. 10.1016/j.canlet.2021.11.020. [DOI] [PubMed] [Google Scholar]
- 33.Nie S, Zhu Y, Yang J, Xin T, Xue S, Sun J, et al. Clinicopathologic analysis of microscopic tumor extension in glioma for external beam radiotherapy planning. BMC Med. 2021;19(1):269. 10.1186/s12916-021-02143-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Yadav N, Mohanty A, Tiwari VA. Fractal dimension and lacunarity measures of glioma subcomponents are discriminative of the grade of gliomas and IDH status. NMR Biomed. 2024;37(12):e5272. 10.1002/nbm.5272. [DOI] [PubMed] [Google Scholar]
- 35.Smits M, van den Bent MJ. Imaging correlates of adult glioma genotypes. Radiology. 2017;284(2):316–31. 10.1148/radiol.2017151930. [DOI] [PubMed] [Google Scholar]
- 36.Park YW, Han K, Ahn SS, Bae S, Choi YS, Chang JH, et al. Prediction of IDH1-Mutation and 1p/19q-Codeletion status using preoperative MR imaging phenotypes in lower grade gliomas. AJNR Am J Neuroradiol. 2018;39(1):37–42. 10.3174/ajnr.A5421. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Yue S, Li S, Huang X, Liu J, Hou X, Zhao Y, et al. Machine learning for the prediction of acute kidney injury in patients with sepsis. J Transl Med. 2022;20(1):215. 10.1186/s12967-022-03364-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Truong NCD, Bangalore Yogananda CG, Wagner BC, Holcomb JM, Reddy D, Saadat N, et al. Two-Stage training framework using multicontrast MRI radiomics for IDH mutation status prediction in glioma. Radiol Artif Intell. 2024;6(4):e230218. 10.1148/ryai.230218. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Kandemirli SG, Kocak B, Naganawa S, Ozturk K, Yip SSF, Chopra S, et al. Machine Learning-Based multiparametric magnetic resonance imaging radiomics for prediction of H3K27M mutation in midline gliomas. World Neurosurg. 2021;151:e78–85. 10.1016/j.wneu.2021.03.135. [DOI] [PubMed] [Google Scholar]
- 40.Yan Z, Wang J, Dong Q, Zhu L, Lin W, Jiang X. XGBoost algorithm and logistic regression to predict the postoperative 5-year outcome in patients with glioma. Ann Transl Med. 2022;10(16):860. 10.21037/atm-22-3384. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets generated or analyzed during the study are not publicly available but are available from the corresponding author on reasonable request.