Machine learning of whole-brain resting-state fMRI signatures for individualized grading of frontal gliomas

Abstract

Purpose

Accurate preoperative grading of gliomas is critical for therapeutic planning and prognostic evaluation. We developed a noninvasive machine learning model leveraging whole-brain resting-state functional magnetic resonance imaging (rs-fMRI) biomarkers to discriminate high-grade (HGGs) and low-grade gliomas (LGGs) in the frontal lobe.

Methods

This retrospective study included 138 patients (78 LGGs, 60 HGGs) with left frontal gliomas. A total of 7134 features were extracted from the mean amplitude of low-frequency fluctuation (mALFF), mean fractional ALFF, mean percentage amplitude of fluctuation (mPerAF), mean regional homogeneity (mReHo) maps and resting-state functional connectivity (RSFC) matrix. Twelve predictive features were selected through Mann-Whitney U test, correlation analysis and least absolute shrinkage and selection operator method. The patients were stratified and randomized into the training and testing datasets with a 7:3 ratio. The logical regression, random forest, support vector machine (SVM) and adaptive boosting algorithms were used to establish models. The model performance was evaluated using area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity.

Results

The selected 12 features included 7 RSFC features, 4 mPerAF features, and 1 mReHo feature. Based on these features, the model was established using the SVM had an optimal performance. The accuracy in the training and testing datasets was 0.957 and 0.727, respectively. The area under the receiver operating characteristic curves was 0.972 and 0.799, respectively.

Conclusions

Our whole-brain rs-fMRI radiomics approach provides an objective tool for preoperative glioma stratification. The biological interpretability of selected features reflects distinct neuroplasticity patterns between LGGs and HGGs, advancing understanding of glioma-network interactions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40644-025-00920-x.

Introduction

Gliomas are the most common primary malignant brain tumors in adults [1] and can be classified as low-grade gliomas (LGGs) and high-grade gliomas (HGGs). This grading can directly reflect the clinical course and prognosis of patients. Despite treatment options including chemotherapy, radiotherapy and surgery, the prognosis of HGGs remains poor [2]. The 5-year relative survival rate after the diagnosis of brain tumors was 35.8%, of which the most aggressive polymorphinoblastoma had the lowest survival rate, 6.8% [3]. Accurate tumor grading can optimize clinical treatment. Therefore, assessment of tumor grading is essential in patients with gliomas. Although gliomas grading remains based on a histopathological approach to tissue sampling, surgical sampling carries greater risk and can lead to permanent neurological impairment or death [4], with the former occurring in up to 8% of patients [5]. Furthermore, biopsy or subtotal resection may not yield representative tumor tissue and may lead to misclassification [6]. Therefore, accurate and reliable noninvasive imaging-based grading of gliomas is desirable. Magnetic resonance imaging (MRI) plays a crucial role in preoperative preparation, surgical planning and assessment of treatment outcomes for intracranial tumors. Some studies on conventional [7–9] and functional MRI [10] of lesions have contributed to identifying gliomas grade. However, they were limited to intergroup comparisons. Thus, individualized gliomas grade assessment can be challenging to perform.

In recent years, machine learning combined with multimodal imaging has been widely used to individually predict gliomas grade [11, 12]. Wu et al. [12] extracted resting-state functional magnetic resonance imaging (rs-fMRI) features in gliomas regions for grading. Results showed that compared with LGGs, HGGs had more complex anatomical morphological and rs-fMRI features in the tumor region. However, most studies require manual or semi-automated outlining of lesions, which is both time-consuming and subjective. The extraction of brain plasticity features on resting-state fMRI is computerized and does not require lesion outlining, which is relatively objective. Moreover, brain plasticity is often coupled with function, which can reveal the functional prognosis of patients to a certain extent. Therefore, this study aimed to combined whole-brain functional plasticity features on rs-fMRI and the machine learning algorithm to establish a predictive model for preoperative frontal gliomas grading, which can assist in the development of individualized clinical treatment plans and provide a reference for functional prognostic prediction.

Materials and methods

This study was approved by the Ethics Committee of our hospital. All participants provided written informed consent.

Study participants

Gliomas classifications were determined using the latest 2021 World Health Organization (WHO) classification of central nervous system tumors [13]. The inclusion criteria were as follows: (1) patients with left frontal LGGs (WHO grade 2) and HGGs (WHO grades 3 and 4) confirmed on pathological examination; (2) those who were aged 18–70 years and right-handed; (3) those with single lesions; (4) those who underwent high-resolution T1-weighted imaging, rs-fMRI, and fluid-attenuated inversion recovery (FLAIR). The exclusion criteria were as follows: (1) patients with midline shift caused by mass effect; (2) those with lesions involving the right and left hemispheres; (3) those with other intracranial abnormalities, such as arachnoid cysts; (4) those who have received relevant treatment prior to MRI; and (5) those with MRI image acquisition issues, such as image artifacts. All images were acquired within 1 week prior to treatment. Based on these criteria, 78 patients with LGGs and 60 with HGGs in the left frontal lobe were retrospectively identified between 2020 and 2024.

MRI acquisition

All patients underwent imaging with a 3T MRI scanner (Siemens Verio MR, Erlangen, Germany) with an 8-channel head coil. The main parameters of high-resolution T1-weighted imaging were as follows: repetition time (TR) = 2000 ms, echo time (TE) = 17 ms, field of view (FOV) = 201 × 230 mm, and matrix = 256 × 168, thickness = 1 mm. The main parameters of T2-weighted FLAIR were as follows: TR = 8000 ms, TE = 102 ms, FOV = 201 × 230 mm, matrix = 256 × 190, and thickness = 6 mm. The main rs-fMRI parameters were as follows: TR = 2000 ms, TE = 35 ms, thickness/gap = 4.0/1.0 mm, FOV = 256 × 256 mm, flip angle = 90°, matrix = 64 × 64, and 240 timepoints.

Preprocessing

Image preprocessing was performed using the RESTplus software (V1.27, http://restfmri.net/forum/restplus) on MATLAB (2018b, MathWorks, Natick, MA, USA). All images were processed using the following steps: First, the original DICOM format was converted to the NIFTI format. Second, the first 10 time points were removed. Third, the slice timing was set. Fourth, head deflection and shift control were facilitated at 1.5° and 1.5 mm, respectively. Fifth, spatial normalization was conducted. In addition, smoothing, de-linear drift, and covariate regression were used to obtain the amplitude of low-frequency fluctuation (ALFF) and fractional ALFF (fALFF) maps. De-linear drifting, covariate regression, and filtering were used to obtain the ReHo map. Smoothing, de-linear drifting, covariate regression, and filtering were used to obtain the percent amplitude of the fluctuation (PerAF) map and resting-state functional connectivity (RSFC). For smoothing, a 4 × 4 × 4 mm smoothing kernel was used to reduce the effects of spatial noise and differences in brain structures between participants. The covariate regression included signal regression of head movements (Friston-24), white matter, and cerebrospinal fluid. For filtering, a range of 0.01–0.08 Hz was used to eliminate the effects of high-frequency signals from respiratory heartbeat and high-frequency noise.

Feature extraction

First, ALFF was measured to reflect the intensity of spontaneous brain activity in the brain regions [14]. The mean amplitude of the low-frequency fluctuation map was calculated by dividing the average of the whole brain within each voxel in the ALFF map. Second, relative to ALFF, measuring fALFF can improve the sensitivity (SEN) and specificity (SPE) of detecting spontaneous brain activity [15]. Third, PerAF can measure the percentage of blood oxygen level-dependent fluctuations relative to the average blood oxygen level-dependent signal intensity at each time point, and can take the average value over the whole time series. PerAF is an effective and more reliable and direct indicator of a promising metric in rs-fMRI studies at the voxel level [16]. Fourth, ReHo can reflect the temporal homogeneity of neural activity [17]. The mean ReHo map was obtained by assigning the average ReHo of the whole brain to each voxel in the ReHo map. According to the automated anatomical labeling (AAL) atlas, the brain was divided into 116 nodes (45 and 26 in each hemisphere and the cerebellum, respectively). Next, the 116 mean values of the corresponding metrics were obtained. Finally, 7134 features, including 6670 RSFC features, 116 mean regional homogeneity features, 116 amplitude of low-frequency fluctuation features, 116 mfALFF features, and 116 mean percentage amplitude of fluctuation features, were used as variables in the following analysis.

Feature selection

In total, 78 patients with LGGs and 60 with HGGs in the left frontal lobe were stratified and randomized into the training and testing datasets with a ratio of 7:3. All steps of feature selection were conducted in the training dataset. Since the number of features was significantly higher than the number of samples, it was at risk for dimensional disasters in high-dimensional cases. Therefore, irrelevant features should be removed to reduce difficulties in the learning task and to improve the learning speed. The predictive features were selected using the following steps: First, significantly different features between the two groups were retained using the Mann–Whitney U test (P < 0.05). Second, to weaken the effect of multicollinearity, variables with strong pairwise correlations must be removed. In this study, variables with a high mean absolute correlation were removed, and the absolute correlation threshold was set at 0.65. Third, based on the results of the first two selection steps, the least absolute shrinkage and selection operator method was applied to retain the most important features. Finally, the retained features were used to build the radiomics prediction models.

Establishment and evaluation of the models

Based on the performance, several classifiers and their learning models were analyzed to determine the optimal model. These classifiers included random forests, logistic regression, support vector machine (SVM), and adaptive boosting. Accuracy (ACC), SEN, SPE, and area under the receiver operating characteristic (ROC) curve (AUC) were used to assess the performance of the prediction model. The equations for these metrics are shown below:

1

2

3

TP, TN, FP, and FN refer to true–positive, true–negative, false–positive, and false-negative findings, respectively. The best model was trained using 10-fold validation in the training dataset. Model building and performance analysis were performed in R 3.5.0 (https://www.r-project.org/) and the “e1071,” “glmnet,” “pROC,” and “caret” package.

Validative analysis

To validate the stability and reproducibility of the model, the effect of other alternative global signal regression strategies and brain segmentation alignments on the main results of this study were evaluated using the same feature extraction, selection methods, and classification model training strategy. Debate regarding the complex composition of the global signal, which likely originates from physiological nuisances (e.g., respiration and movement) and neuronal signals, is still ongoing. Therefore, a complementary analysis of global signal regression was performed to determine if global signal regression is required for data processing. To identify if the results of the model based on the AAL atlas could be generalized to other analyses or models specific to certain analyses, another analysis was conducted using a functional analysis atlas comprising 160 nodes (Dos-160 atlas) [18]. The SVM model with global signal regression and the SVM model using the Dos-160 atlas were obtained and applied in the testing dataset.

Comparative analysis

To compare with T1-weighted enhancement-based model, the same feature extraction, selection methods, and classification model training strategy were performed for the region of interests (ROIs). The ROIs including tumor parency, intratumoral necrosis, hemorrhage, cystic degeneration and peritumoral edema were delineated based on T1-weighted enhanced and T2-weighted FLAIR images using 3DSlicer software (a free and open source toolkit, version 4.10.2, available at: https://download.slicer.org/).

Statistical analysis

Age and lesion volume were compared between the two groups using the two-sample t-test. Sex was compared using the chi-square test, and disease duration was compared with the Mann–Whitney U test. These variables were statistically compared using the Statistical Package for the Social Sciences software (version 26.0). Quantitative data were presented as mean ± standard deviation. A P value of < 0.05 was considered statistically significant difference. Figure 1 shows the general flow of this study.

Fig. 1.

Flow chart of establishing the prediction model

Results

Demographic characteristics of the participants

Table 1 shows the clinical demographic characteristics of the patients. The two groups differed in terms of age, sex, and lesion volume (P < 0.05). In this cohort, Patients with HGGs were more likely to be men and older and had a larger lesion volume than those with LGGs. There was no significant difference in the disease course between the two groups (P > 0.05). In terms of clinical symptoms, patients with LGGs most commonly presented with epilepsy. Meanwhile, patients with HGGs typically had more heterogenous symptoms.

Table 1.

Demographic characteristics of two groups

Variable	LGGs(n = 78)	HGGs(n = 60)	P value
Age (years)	36.98 ± 10.01	47.75 ± 14.42	0.002
Sex (male/female)	29/49	33/27	0.037
Disease duration (months)
Median	1.0	1.0	0.648
IQR	1.625	4.75
Range	0.25-36	0.5–36
Lesion volume (cm 3 )	20.54 ± 14.85	58.01 ± 41.56	0.000
Symptoms
Epilepsy	44	15
Asymptomatic	12	9
Headache	10	11
Dizzy	5	9
Limb weakness	3	5
Numbness of limbs	3	6
Aphasia	1	5
Histology
Astrocytoma	57	5
Oligodendroglioma	21	10
Glioblastoma		45
IDH status
Mutant	78	15
Wildtype		45

HGGs, high-grade gliomas; IDH, isocitrate dehydrogenase; IQR, interquartile range; LGGs, low-grade gliomas. Quantitative variables are expressed as the mean ± standard deviation

Feature selection

Among the 7134 features, 4361 were retained using the Mann–Whitney U test. Further, 857 features were retained by removing variables with a strong correlation, and 12 features were retained using the least absolute shrinkage and selection operator method. The features included 7 RSFC features, 4 PerAF features, and 1 ReHo feature. Finally, these features were used to build the prediction model. The RSFC features were located in the DMN, affective network (AN), and sensorimotor network (SMN) (Table 2; Fig. 2). The other features included PerAF of the left superior frontal gyrus, medial orbital (ORBsupmed), PerAF of the left insula (INS), PerAF of the left inferior parietal, supramarginal, and angular gyri (IPL), PerAF of the left superior temporal gyrus (STG), and ReHo of the right Heschl’s gyrus (HES.R). In addition, to validate whether the selected feature values increased or decreased between the two groups, the mean and standard deviation of the selected features were calculated in both groups (Table 3). Figure 3 shows the importance of these characteristics in establishing the model.

Table 2.

The RSFC features for discriminating patients with LGGs and HGGs

AAL	Brain area	Network	AAL	Brain area	Network
2	PreCG.R	SMN	18	ROL.R	SMN
7	MFG.L	DMN	63	SMG.L	DMN
23	SFGmed.L	DMN	24	SFGmed.R	DMN
26	ORBsupmed.R	DMN	32	ACG.R	DMN
37	HIP.L	DMN	42	AMYG.R	AN
41	AMYG.L	AN	42	AMYG.R	AN
65	ANG.L	DMN	87	TPOmid.L	AN

AN, affective network; DMN, default mode network; SMN, sensorimotor network; PreCG, Precental gyrus; MFG, Middle frontal gyrus; SFGmed, Superior frontal gyrus, medial; ORBsupmed, Superior frontal gyrus, medial orbital; HIP, Hippocampus; AMYG, Amygdala; ANG, Angular gyrus; ROL, Rolandic operculum; SMG, Supramarginal gyrus; ACG, Anterior cingulate and paracingulate gyri; TPOmid, Temporal pole: middle temporal gyrus; L, left; R, right

Fig. 2.

Seven RSFC features used to differentiate LGGs from HGGs. HGGs, high-grade gliomas; LGGs, low-grade gliomas; RSFC, resting-state functional connectivity

Table 3.

Characteristic of the 12 features for discriminating patients with LGGs and HGGs

Features	LGGs(n = 78)	HGGs(n = 60)	P值
PreCG.R-ROL.R	0.963 ± 0.213	0.974 ± 0.257	0.037
MFG.L-SMG.L	1.016 ± 0.361	0.675 ± 0.459	0.001
SFGmed.L-SFGmed.R	0.916 ± 0.200	1.071 ± 0.196	0.003
ORBsupmed.R-ACG.R	1.026 ± 0.249	0.829 ± 0.244	0.003
HIP.L-AMYG.R	1.052 ± 0.361	0.624 ± 0.365	0.000
AMYG.L-AMYG.R	1.043 ± 0.278	0.697 ± 0.415	0.001
ANG.L-TPOmid.L	0.982 ± 0.454	0.622 ± 0.523	0.004
mPerAF of ORBsupmed.L	0.791 ± 0.433	1.071 ± 0.419	0.012
mPerAF of INS.L	-0.898 ± 0.213	-1.083 ± 0.264	0.002
mPerAF of IPL.L	-0.485 ± 0.559	-1.090 ± 0.811	0.001
mPerAF of STG.L	-0.426 ± 0.418	-1.256 ± 0.787	0.000
mReHo of HES.R	-0.499 ± 0.933	-0.103 ± 0.888	0.002

INS, insula; IPL, Inferior parietal, but supramarginal and angular gyri; STG, Superior temporal gyrus; HES, Heschl gyrus; LGGs, low-grade gliomas; HGGs, high-grade gliomas;.L, left;.R, right

Fig. 3.

Importance of the 12 features used to build predictive models

Model and its performance

With the 12 features, the prediction model in the training dataset was built using various classifier algorithms. Finally, the optimal model was obtained using the SVM algorithm, with an ACC of 0.957 and an AUC of 0.972 in the training dataset. In the testing dataset, the model’s ACC was 0.727, and the AUC was 0.799 (Fig. 4; Table 4). Results showed that the selected 12 features can be used to predict HGGs and LGGs.

Fig. 4.

ROC curves of the SVM model predicting high- and low-grade gliomas in the training (A) and testing (B) datasets, respectively. ROC, receiver operating characteristic curve; SVM, support vector machine

Table 4.

The results of the major work and validation analysis

	Training set				Testing set
	AUC	ACC	SEN	SPE	AUC	ACC	SEN	SPE
Model 1	0.972	0.957	1	0.875	0.799	0.727	0.786	0.75
Model 2	0.972	0.957	0.935	1	0.826	0.682	0.786	0.875
Model 3	0.982	0.915	0.871	1	0.714	0.682	0.571	0.875

Model 1, AAL 116 atlas without regression of global signal; Model 2, AAL 116 atlas with regression of global signal; Model 3, Dos-160 atlas without regression of global signal

Validative analysis

The ACC of the SVM model established after preprocessing with the global signal regression were 0.957 with an AUC of 0.972 in the training dataset, and 0.682 with an AUC of 0.826 in the testing dataset (Table 4). The performance of the model constructed using the data with the global signal regression in the training dataset was similar to that of the model developed without the global signal regression. However, its performance in the testing dataset was poor. Therefore, the model constructed using global signal regression processing could be at risk of overfitting. Therefore, the models without the global signal regression could be more stable.

The ACC of the SVM model built by extracting features using the Dos-160 atlas were 0.915 with an AUC of 0.982 in the training dataset, and 0.682 with an AUC of 0.714 in the testing dataset. The results obtained using the Dos-160 atlas were similar to those obtained using the structural AAL atlas. Thus, our results can be generalized to other segmentation atlases. However, the structural AAL atlas reflects the physical anatomy of the brain, and the functional Dos-160 atlas reflects the functional activation regions of the brain in the time series. These two components could interpret the brain from different dimensions. Therefore, the obtained features and results might not be completely consistent (Table 4).

Comparative analysis

A total of 1195 features were extracted from T1-weighted enhanced images of ROIs. 10 optimized radiomics features were finally selected to establish models (Supplementary Fig. 1). Similar to rs-fMRI, the optimal model was obtained using the SVM algorithm, with an ACC of 0.935, a SEN of 0.971, a SPE of 0.907 and an AUC of 0.970 in the training dataset. In the testing dataset, the model’s ACC was 0.794, the SEN was 0.789, the SPE was 0.800, and the AUC was 0.850 (Supplementary Fig. 2).

Discussion

Gliomas grading is closely related to patient prognosis and is important for clinical decision-making. In this study, whole-brain functional activity (ALFF, fALFF, PerAF, ReHo maps) and connectivity features (RSFC) were extracted based on rs-fMRI for building a prediction model of high- and low- grade of frontal gliomas. The optimal prediction model was obtained by SVM algorithm, with ACC of 0.957 and 0.727 in the training and testing datasets, respectively.

In the cohort of this study, patients with HGGs were significantly older and had larger lesion volume than patients with LGGs, and HGGs were more often seen in male patients, which are consistent with the clinical characteristics of gliomas [2, 19]. Meanwhile, epilepsy was most common in patients with LGGs. This may be related to the persistent abnormal discharges that result from slower progression of LGGs and allow epileptogenic mechanisms to develop in parallel with neuroplasticity [20]. However, rapid, extensive, and severe brain tissue destruction limits brain plasticity and the spread of abnormal synchrony in patients with HGGs; thus, reducing the incidence of epilepsy and leading to a more heterogeneity of clinical symptoms. This also supports the finding that the incidence of glioma-related epilepsy gradually decreases with the increase of tumor malignancy [21].

rs-fMRI has potential as an adjunctive clinical diagnostic tool [12]. The selected features based on rs-fMRI included seven RSFC features and five brain functional activity features. The selected RSFC features were distributed in the DMN, AN, and SMN. The DMN is the center of spontaneous cognition, self-referential processing, and emotion regulation [22]. An abnormal DMN can cause self-referential and affective processing dysfunction, which manifests as excessive negative self-focus [23]. Esposito et al. [24] have found that gliomas can cause changes in DMN connectivity and differ between patients with LGGs and those with HGGs. Harris et al. [10] revealed that patients with HGGs had a lower DMN integrity than those with LGGs. In the current study, most RSFC features were located in the DMN, and patients with HGGs had a lower RSFC than those with LGGs. This can be related to a combination of factors, including true-positive changes caused by increased tumor infiltration and disrupted functional reorganization, and false-negative signals attributed to tumor-induced neurovascular uncoupling and, possibly, hemodynamic alterations that artificially reduce the RSFC [25, 26]. The AN is dedicated to emotion processing and regulation [27]. Structural and functional abnormalities in the AN can lead to mood dysregulation. Brain region of RSFC features located in the AN are predominantly amygdala, and the connectivity feature between the left hippocampus and right amygdala is the most important contributor of all the retained features. Recently, patients with LGGs have a significantly larger gray matter volume in the contralesional amygdala [28]. These findings may be related to the increased and enhanced functional connectivity of other brain regions with the amygdala and may be involved in functional impairment or compensation. Therefore, the amygdala has an important role in functional impairment or compensation in patients with gliomas. In addition, structural and functional alterations in the amygdala can be related to gliomas malignancy. The SMN is important for primary sensory and motor cortical functions. Recently, based on the graph theoretical analysis of the RSFC matrix, Fang et al. [29] found that the contralesional SMN is involved in motor function compensation in patients with gliomas. In addition, glioma-related epilepsy (GRE) is the most common symptom of prefrontal gliomas [30]. In this cohort, epilepsy was the most common symptom in patients with LGGs, and patients with HGGs had more heterogeneous symptoms. Fang et al. [31] examined alterations in the functional network of patients with non-GRE and those with GRE from prefrontal gliomas. All significant alterations were present only in the SMN. Therefore, the connectivity of the precentral gyrus and Rolandic operculum located in the SMN of the contralesional hemisphere may reflect the patient’s motor function compensation and epileptic symptoms. These studies, combined with the aberrant functional connectivity in the DMN, AN, and SMN, may indicate that the balance between network integration and segregation is disrupted to various degrees in patients with HGGs and LGGs.

In terms of features of functional activity for the establishment of the model, in patients with gliomas, INS located in the salience network has increased in ALFF and is correlated with tumor grade [32]. The salience network modulates the activation of the DMN and central executive network by detecting the presence of salient stimuli [33]. The STG comprises functional subregions that play an important role in the cerebral cortex, and it is closely associated with speech and language processing [34, 35]. Further, it undergoes structural and functional alterations in different psychiatric and psychological disorders [36, 37]. The HES, a subregion of the STG, is located in the superior temporal plane and is essential for the initial processing of auditory information [38] and is also involved in emotional and memory processing [39]. The significant increase in the gray matter volume of contralesional STG in patients with frontal LGGs has been well established [28]. The results of this study support the notion that gliomas induce changes in brain activity in these brain regions, which differ between patients with HGGs and LGGs.

Notably, our comparative analysis revealed the T1-weighted enhancement-based SVM model achieved marginally higher performance in the testing cohort (ACC 0.794 vs. 0.727, AUC 0.850 vs. 0.799). This finding may reflect fundamental differences in captured pathophysiological processes. T1-weighted enhancement primarily detects blood-brain barrier breakdown [9], whereas rs-fMRI biomarkers map tumor-induced functional network reorganization. The combination of structural and functional biomarkers may create synergistic diagnostic value - a promising direction for future multimodal integration. When compared with previous MRI-based grading studies, our rs-fMRI based model offers distinct advantages in probing tumor-network interactions. Conventional approaches using perfusion-weighted imaging or diffusion tensor imaging primarily assess intratumoral heterogeneity. In contrast, our whole-brain rs-fMRI paradigm captures both local activity alterations and global network disruptions - features particularly relevant for frontal gliomas that frequently infiltrate functional networks. Importantly, our automated feature selection pipeline overcomes the manual ROI-delineation requirements of traditional MRI radiomics, potentially enhancing clinical translation feasibility.

Overall, alterations in whole-brain functional activity and connectivity may reflect the functional imaging markers of HGGs and LGGs. These findings also contribute to a better understanding of the biological behaviors underlying HGGs and LGGs. The current study has novel findings, and the model has a better performance in predicting gliomas grade. However, more studies should be conducted to validate and improve the reliability, accuracy, and reproducibility of the model. The current study had several limitations. First, it collected data from a single center, lacked external validation datasets, and included a limited sample size, especially for patients with HGGs. Second, it only classified gliomas into HGGs and LGGs. Thus, subsequent studies could further subdivide the grades of gliomas and perform grade prediction (e.g., WHO grades 2, 3, and 4). Third, due to the retrospective nature of this study, there was a lack of relevant functional scale assessments. Furthermore, right hemisphere gliomas had a contralesional higher DMN-RSFC than left hemisphere gliomas [40]. However, only the predictive analysis of left frontal gliomas was performed in this study. Therefore, gliomas in other sites needs to be further explored. Finally, preprocessing, parameter calculations and validative analysis performed in our study can mitigate variability to a certain extent, while the translational potential of rs-fMRI biomarkers still necessitates careful consideration of measurement reproducibility. Emerging harmonization techniques such as ComBat (a batch-effect correction tool) have shown promise in reducing cross-sectional and longitudinal variability [41], suggesting an important direction for future validation studies.

Conclusion

This radiomics method based on whole-brain functional activity and connectivity can individually identify HGGs and LGGs with a high accuracy. Hence, it can be a potential adjunct to clinical diagnostic systems. In the future, more information, such as clinical function assessments, should be included in the classification model, which can provide clinicians and researchers with important insights on areas with a higher diagnostic accuracy, individualized treatment, and prognostic guidance. Meanwhile, this study can be a reference for functional prognostic assessment and the prediction of patients with brain tumors.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Guarantor of integrity of the entire study, study concepts and design (DYG and KL); literature research, clinical studies, experimental studies / data analysis and statistical analysis (YH, XC and HYC); manuscript preparation and editing (KL and HY). All authors contributed to this article and approved the final version.

Funding

This work was supported by the National Natural Science Foundation of China [grant numbers 82372048], Research Startup Fund of Huashan Hospital, Fudan University [grant numbers 2021QD035], Shanghai Sailing Program [grant numbers 22YF1405000], Shanghai Municipal Commission of Science and Technology [grant numbers 22TS1400900, 23S31904100, 22ZR1409500 and 22S31905300], Greater Bay Area Institute of Precision Medicine (Guangzhou) [grant numbers KCH2310094], and Science and Technology Commission of Shanghai Municipality [grant numbers 24SF1904200, 24SF1904201].

Data availability

The data and materials of this study are available from the corresponding author on reasonable request.

Declarations

Human ethics and consent to participate

All procedures performed in the studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. This study was approved by the institutional review board of our hospital. Informed consent was obtained from all individual participants included in the study.

Competing interests

The authors declare no competing interests.