IMR Press / JIN / Volume 23 / Issue 5 / DOI: 10.31083/j.jin2305100
Open Access Systematic Review
The Current Diagnostic Performance of MRI-Based Radiomics for Glioma Grading: A Meta-Analysis
Show Less
1 Division of Neurosurgery, Department of Surgical Specialties, Radiological Sciences, and Public Health, University of Brescia, 25123 Brescia, Italy
2 Interuniversity Department of Regional and Urban Studies and Planning, Politecnico di Torino, 10125 Torino, Italy
3 Department of Medical Artificial Intelligence, Konyang University, 35365 Daejeon, Republic of Korea
4 Department of Radiology and Nuclear Medicine, University of Oslo, 0372 Oslo, Norway
5 Department of Medical Physics, University of Thessaly, 413 34 Larissa, Greece
6 Department of Mathematical Sciences “Giuseppe Luigi Lagrange”, Politecnico di Torino, 10123 Torino, Italy
7 Department of Ophthalmology, University Hospital of Udine, 33100 Udine, Italy
8 Neurosurgery Unit, Head-Neck and NeuroScience Department University Hospital of Udine, p.le S. Maria della Misericordia 15, 33100 Udine, Italy
9 Department of Advanced Biomedical Sciences, University of Naples “Federico II”, 80126 Naples, Italy
10 Department of Neurosurgery and Interventional Neuroradiology, Mayo Clinic, Rochester, MN 55905, USA
*Correspondence: markzeppieri@hotmail.com (Marco Zeppieri)
J. Integr. Neurosci. 2024, 23(5), 100; https://doi.org/10.31083/j.jin2305100
Submitted: 13 November 2023 | Revised: 28 December 2023 | Accepted: 4 January 2024 | Published: 14 May 2024
Copyright: © 2024 The Author(s). Published by IMR Press.
This is an open access article under the CC BY 4.0 license.
Abstract

Background: Multiple radiomics models have been proposed for grading glioma using different algorithms, features, and sequences of magnetic resonance imaging. The research seeks to assess the present overall performance of radiomics for grading glioma. Methods: A systematic literature review of the databases Ovid MEDLINE PubMed, and Ovid EMBASE for publications published on radiomics for glioma grading between 2012 and 2023 was performed. The systematic review was carried out following the criteria of Preferred Reporting Items for Systematic Reviews and Meta-Analysis. Results: In the meta-analysis, a total of 7654 patients from 40 articles, were assessed. R-package mada was used for modeling the joint estimates of specificity (SPE) and sensitivity (SEN). Pooled event rates across studies were performed with a random-effects meta-analysis. The heterogeneity of SPE and SEN were based on the χ2 test. Overall values for SPE and SEN in the differentiation between high-grade gliomas (HGGs) and low-grade gliomas (LGGs) were 84% and 91%, respectively. With regards to the discrimination between World Health Organization (WHO) grade 4 and WHO grade 3, the overall SPE was 81% and the SEN was 89%. The modern non-linear classifiers showed a better trend, whereas textural features tend to be the best-performing (29%) and the most used. Conclusions: Our findings confirm that present radiomics’ diagnostic performance for glioma grading is superior in terms of SEN and SPE for the HGGs vs. LGGs discrimination task when compared to the WHO grade 4 vs. 3 task.

Keywords
glioma grading
radiomics
magnetic resonance imaging (MRI) features
systematic review
meta-analysis
1. Introduction

With the 5th edition of the World Health Organization (WHO) Classification of Tumors of the Central Nervous System (CNS) published in 2021 [1], a major role has been assigned to molecular patterns for the differential diagnosis of gliomas [2]. These innovations have been followed by improvements in therapeutic strategies, with more targeted and focused therapies [3]. The gold standard for diagnosis of gliomas remains biopsy, although it puts the patients at inevitable risk of procedure-related complications [4]. Over the last 10 years, artificial intelligence (AI) applied to diagnostic imaging progressively gained more popularity [5]. Radiomics, through deep learning (DL) and machine learning (ML) techniques, refers to the extraction of mineable data from medical imaging, boosting its diagnostic capability. Since its first appearance in 2012 as a computer-aided detection and diagnosis system [6], radiomics has grown massively infiltrating the diverse fields of neuro-oncology [4]. Clinical applications of these noninvasive radiomics-based models range from screening and characterization to monitoring and prediction [7]. This has clinical implications for any CNS tumor, from gliomas to meningioma and pituitary tumors [8, 9, 10].

The grading of glioma has been one of the main focuses of radiomics [11]. Nowadays, multiple radiomics-based models have been proposed for grading glioma, using different features, models, and sequences of magnetic resonance imaging (MRI). Each carried out its diagnostic accuracy, thus, the actual performance of radiomics is dependent on the specific model and is not standardized [12]. Given the heterogeneity and rapid expansion of this technique, we aimed to better define its overall potential in grading gliomas. Therefore, based on the current literature we conducted a meta-analysis and systematic review to evaluate the current performance state of radiomics for the grading of brain gliomas.

2. Materials and Methods
2.1 Systematic Review and Inclusion Criteria

The Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) criteria were followed in conducting the systematic review [13]. An expert librarian created and carried out a thorough literature search of the Ovid MEDLINE, PubMed, and Ovid EMBASE databases with advice from the authors. Three terms were utilized in “AND” combinations: “radiomics”, “glioma”, and “grading”. Only publications published between 2012 and 2023 were included in the search.

Two Authors (L.D.M. and F.P.) assessed the study inclusion criteria during the review process. The following inclusion criteria included: (1) Written in the language of English; (2) case series including based on 10 patients or more; (3) studies reporting exclusively histologically proven brain gliomas; (4) studies that included the WHO grade; (5) studies based on radiomics; (6) studies that reported the performance data of radiomics for glioma grade prediction.

2.2 Data Abstraction

Three investigators (L.D.M., F.P., and E.A) independently collected data from systematic reviews that were eligible using piloted forms. Each systematic review was then validated by a different investigator, and consensus was used when confronted with disagreements by. Study characteristics were obtained from each eligible systematic, based on the baseline information that included: year of publication, number of patients, WHO grade, and MRI protocol. As for the radiomics model, we collected information about: the AI sub-category (i.e., ML or DL), the classification algorithm (i.e., logistic regression (LR), gradient boost (GB), naïve bayes (NB), support vector machine (SVM), multilayer perceptron (MLP), elastic net regression (ENR), linear discriminant analysis (LDA), nearest neighbors (NN), random forest (RF), convolutional neural network (CNN), and others), the selected features (i.e., textural, geometrical or morphological, voxel intensities-based, others), the employed MRI modalities (i.e., T1-weighted [T1W], T2-weighted [T2W], others), and the application of cross-validation analysis (i.e., yes or no).

2.3 Quality Assessment

We analyzed the clarity of reporting in eligible systematic reviews with the PRISMA checklist. The checklist includes 27 items that are used to assess the reporting quality. The PRISMA checklist is shown in Supplementary Material [13].

2.4 Outcomes

Our primary outcomes were the specificity (SPE), sensitivity (SEN), and Summary Receiver Operating Characteristics (SROC) curve of radiomics for predicting the WHO grade of gliomas of the brain. Bivariate analyses by discrimination task between low-grade gliomas (LGGs; WHO grade 1 and 2) vs. high-grade gliomas (HGGs; WHO grade 3 and 4), and WHO grade 4 gliomas vs. WHO grade 3 gliomas were conducted.

The impact of the following variables on the performance of the proposed radiomics models was evaluated as secondary outcomes: year of development, cohort size, AI sub-category, validated classifiers, selected features, MRI modalities, and cross-validation strategy. These variables were also studied in quantitative terms to picture the current trends of radiomics models for glioma grading.

2.5 Study Risk of Bias Assessment

To evaluate the methodologic quality of the studies that were part of our meta-analysis, we adjusted the Newcastle-Ottawa Scale (NOS) [14]. The purpose of this tool is to be used in comparative investigations. However, as our investigations lacked a control group, we evaluated the methodologic quality of the data using a subset of the scale’s items, paying particular attention to the following queries: (1) Was a random sample used in the study, or did all patients or consecutive patients participate? (2) Was the research prospective or retrospective? (3) Did the clinical follow-up provide enough information to determine every outcome? (4) Were the results published? (5) Were the criteria for inclusion and exclusion well-defined [15]?

2.6 Statistical Analysis

Data from primary studies were reported in a 2 × 2 contingency table consisting of true positive (TP), false negative (FN), false positive (FP), and true negative (TN) based on the concordance between biopsy results and the radiomics tool predictions. Such a table served as input for the R-package mada (https://cran.r-project.org/web/packages/mada/index.html) [16], used for modeling the joint estimates of SEN and SPE and their 95% confidence intervals (CIs). A random-effects meta-analysis was used to pool the event rates across studies, and the χ2 test was performed to analyze the heterogeneity of SPE and SEN, considering the null hypothesis as equality in each case.

3. Results
3.1 Literature Reviews

After duplicates were eliminated, 390 papers were found. 103 articles were found for full-text study following the analysis of the titles and abstracts. Forty papers were evaluated for eligibility. The following criteria led to the exclusion of the remaining 63 articles: (1) studies not reporting data on radiomics performance for glioma grading (34 articles), (2) studies not reporting the WHO grade (13 articles), (3) studies reporting on AI-based models other than radiomics (8 articles), (4) improper study design (5 articles), and (5) language other than English (3 articles). The Authors of three included articles [17, 18, 19] provided missing performance data and the results were integrated into the data abstraction process. For each of the patient groups under consideration, at least one or more outcome measures were available for all of the studies that were part of the analysis. The PRISMA statement’s flow chart is seen in Fig. 1.

Fig. 1.

PRISMA flow diagram showing the search process of the literature. Abbreviations: PRISMA, preferred reporting items for systematic reviews and meta-analysis; WHO, World Health Organization.

3.2 Baseline and Radiomics Data

A total of 7654 patients were included in this study. Most studies were published in 2022 (25%), followed by 2021 (20%), and 2020 (17.5%). The smallest study included 26 patients, while the largest 572. Differentiation in LGG and HGG was reported for 6951 patients (90.8%), of which 4537 had HGGs (65.3%). Among those with HGGs, discrimination between WHO grade 4 and WHO grade 3 was reported for 2742 patients (60.4%), of which 1781 had grade 4 gliomas (65%). Each study included different MRI sequences and the most common was the contrast-enhanced T1W (CE-T1W; 92.5%), followed by T2W (T2W; 77.5%), T2W fluid-attenuated inversion recovery (T2-FLAIR; 70%), and T1W (65%). Other MRI modalities were diffusion-weighted imaging (DWI; 20%), perfusion-weighted imaging (PWI; 15%), and proton magnetic resonance spectroscopy (1H-MRS; 5%).

A number of 31 studies reported on ML (77.5%), 3 articles reported on DL (7.5%), and 6 were hybrid studies (15%). As for the classifiers, LR was the most adopted (55%) followed by SVM (40%), RF (30%), CNN (17.5%), GB (7.5%), and others. Table 1 (Ref. [17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56]) shows a summary of the studies included.

Table 1.Summary of studies included.
Author, Journal, Year Patients (n.) WHO Grade (n.) MRI Protocol Method Classifiers
LGG HGG
3 4
Chen, Int J Biomed Imaging, 2018 [20] 274 54 220 NA NA T1W, CE-T1W, T2W, T2-FLAIR ML & DL SVM, CNN
Cheng, IEEE J Biomed Health Inform, 2022 [21] 438 119 319 NA NA T1W, CE-T1W, T2W, T2-FLAIR DL CNN
Cheng, IEEE/ACM Trans Comput Biol Bioinform, 2022 [22] 350 92 258 NA NA T1W, CE-T1W, T2W, T2-FLAIR ML LR, SVM, RF, GB
Cho, Annu Int Conf IEEE Eng Med Biol Soc, 2017 [23] 108 54 54 NA NA T1W, CE-T1W, T2W, T2-FLAIR ML LR
Cho, PeerJ, 2018 [19] 285 75 210 NA NA T1W, CE-T1W, T2W, T2-FLAIR ML LR, SVM, RF
Ding, Quant Imaging Med Surg, 2022 [24] 50 NA NA NA 25 CE-T1W ML & DL RF, CNN
Ditmer, J Neurooncol, 2018 [25] 94 14 80 NA NA CE-T1W ML LR
Gao, Front Oncol, 2020 [26] 369 147 222 116 106 CE-T1W ML LR, SVM, RF
Gihr, Front Oncol, 2020 [27] 26 26 0 0 0 T2W, DWI ML LR
Guo, Diagn Interv Radiol, 2021 [28] 152 47 105 39 66 T1W, CE-T1W, T2W, T2-FLAIR ML LR
Gutta, Am J Neuroradiol, 2021 [29] 220 59 161 46 115 T1W, CE-T1W, T2W, T2-FLAIR DL CNN
Hashido, J Comput Assist Tomogr, 2021 [30] 52 18 34 NA NA T1W, CE-T1W, T2W, T2-FLAIR, DWI, PWI ML LR, SVM, RF
Hashido, Sci Rep, 2020 [31] 46 15 31 4 27 T1W, CE-T1W, T2W, T2-FLAIR, PWI ML LR
Hu, Comput Biol Med, 2021 [32] 505 233 272 107 165 T1W, CE-T1W, T2W, T2-FLAIR ML SVM
Huang, J Comput Assist Tomogr, 2021 [33] 59 13 46 NA NA T1W, CE-T1W, T2W, T2-FLAIR ML LR
Kobayashi, Sci Rep, 2021 [34] 355 NA NA NA 259 T1W, CE-T1W, T2W, T2-FLAIR ML & DL LR, CNN
Li, Cancers, 2022 [35] 212 105 107 0 107 T1W, CE-T1W, T2W, T2-FLAIR ML SVM
Lin, Med Phys, 2022 [36] 100 50 50 NA NA T1W, CE-T1W, T2W, DWI, 1H-MRS ML LR
Liu, Neuroradiology, 2022 [37] 182 63 119 NA NA T1W, CE-T1W, T2W, T2-FLAIR, DWI ML LR
Lu, Clin Cancer Res, 2018 [38] 214 NA NA NA 106 CE-T1W, T2-FLAIR, DWI ML SVM
Nakamoto, Sci Rep, 2020 [39] 157 NA 157 55 102 CE-T1W, T2W ML LR, SVM, RF, NB, NN
Ning, Ann Transl Med, 2021 [40] 334 112 222 117 105 CE-T1W, T2-FLAIR ML & DL SVM, CNN
Park, Korean J Radiol, 2019 [41] 314 213 101 101 NA CE-T1W, T2W, T2-FLAIR ML RF, GB, ENR, LDA
Reza, J Med Imaging, 2019 [42] 285 75 210 NA NA CE-T1W, T2W, T2-FLAIR ML SVM, RF, GB
Skogen, European Journal of Radiology, 2016 [17] 95 27 68 34 34 T1W, CE-T1W, T2W, T2-FLAIR ML LR
Su, Am J Transl Res, 2021 [43] 139 69 70 36 34 T1W, CE-T1W, T2W, T2-FLAIR, DWI ML LR
Su, Eur Radiol, 2019 [44] 217 95 122 61 61 T1W, CE-T1W, T2W, T2-FLAIR, DWI, PWI ML LR
Sudre, BMC Med Inform Decis Mak, 2020 [45] 333 101 232 74 158 PWI ML RF
Takahashi, Int J Radiat Oncol Biol Phys, 2019 [46] 55 14 41 12 29 T1W, CE-T1W, T2W, T2-FLAIR, DWI ML LR, SVM
Tian, J Magn Reson Imaging, 2018 [47] 153 42 111 33 78 T1W, CE-T1W, T2W, DWI, PWI ML SVM
Vamvakas, Phys Med, 2019 [18] 40 20 20 NA NA T1W, CE-T1W, T2W, T2-FLAIR, DWI, PWI, 1H-MRS ML SVM
van der Voort, Neuro Oncol, 2023 [48] 238 47 191 59 132 T1W, CE-T1W, T2W, T2-FLAIR DL CNN
Wang, J Magn Reson Imaging, 2019 [49] 85 34 51 NA NA CE-T1W, T2W, DWI ML LR
Xie, J Magn Reson Imaging, 2018 [50] 42 15 27 13 14 T1W, CE-T1W, T2W, T2-FLAIR ML NA
Xu, Front Oncol, 2022 [51] 572 190 382 NA NA T1W, CE-T1W, T2W ML & DL RF
Xu, Quant Imaging Med Surg, 2022 [52] 129 62 67 NA NA CE-T1W, T2 FLAIR, DWI ML LR, SVM, RF, NB, NN
Zhang, J Digit Imaging, 2020 [53] 108 43 65 NA NA DWI ML & DL SVM
Zhao, BMC Neurol, 2020 [54] 69 36 33 33 NA CE-T1W, T2-FLAIR ML RF
Zhou, Int J Clin Pract, 2022 [55] 114 35 79 21 58 T1W, CE-T1W, T2W ML LR
Zhou, Neuro Oncol, 2017 [56] 84 NA NA NA 0 T1W, CE-T1W, T2W, T2-FLAIR ML LR

Abbreviations: NA, not available; WHO, world health organization; LGG, low-grade glioma; HGG, high-grade glioma; MRI, magnetic resonance imaging; T1W, T1-weighted; CE-T1W, contrast-enhanced T1W; T2W, T2-weighted; T2-FLAIR, T2W-fluid-attenuated inversion recover; DWI, diffusion-weighted imaging; PWI, perfusion-weighted imaging; 1H-MRS, proton magnetic resonance spectroscopy; ML, machine learning; DL, deep learning; SVM, support vector machine; CNN, convolutional neural network; LR, logistic regression; RF, random forest; GB, gradient boost; NB, naïve bayes; NN, nearest neighbors; ENR, elastic net regression; LDA, linear discriminant analysis.

3.3 Primary Outcomes

The performance of radiomics for LGGs vs. HGGs and WHO grade 3 vs. 4 categorizations was reported for a total of 3290 patients and 704 patients, respectively. Overall SPE and SEN for differentiation between LGGs and HGGs were 91% (95% CI = 0.86–0.94) and 84% (95% CI = 0.78–0.89), respectively. With regards to the discrimination task between WHO grade 4 and WHO grade 3, the overall SEN was 89% (95% CI = 0.82–0.94) and the overall SPE was 81% (95% CI = 0.66–0.91), respectively. Fig. 2 shows the univariate forest plots of the analysis for discrimination between LGGs vs. HGGs and grade 3 vs. grade 4 gliomas. Fig. 3 provides the respective SROC curves for the different binary categorization tasks considered in our analysis. In the canonical receiver operating characteristic (ROC) curve, each data point belongs to a single study in which several different diagnostic thresholds are employed to categorize between two classes of interest (e.g., cases and non-cases). In a single study, changing the threshold results indeed in the plot of the true positive rate (TPR) against the false positive rate (FPR) at each threshold setting. Conversely, in a meta-analysis, the units of analysis are separate and single studies. Hence, the SROC curve aims to represent the relationship between TPR and FPR across studies, recognizing they may have used diverse thresholds. As above mentioned, to fit such curves, we put into effect the R package mada, which, besides the SROC, provides two further figures of merit: (i) the confidence region of the summary estimate, which provides a measure of the ROC-based performance heterogeneity among the included studies; (ii) the prediction region for a new hypothetical study, which is the zone in the false positive rate vs. sensitivity space, in which it is likely to expect a new study.

Fig. 2.

Forest plots for discrimination between LGGs vs. HGGs (A) and grade 3 vs. grade 4 gliomas (B). The red diamonds represent the overall results.

Fig. 3.

SROC curves for the differentiation between LGGs vs. HGGs (A) and grade 3 vs. grade 4 gliomas (B). Each curve provides: (i) the summary estimate of the considered study (see the red circle); (ii) the confidence region of the summary estimate (see the red ellipse-like trace), which furnishes a measure of the performance heterogeneity among the included studies; (iii) the prediction region for a new hypothetical study, where it is likely to expect a new study (see the dashed line). Abbreviations: SROC, summary receiver operating characteristic.

3.4 Secondary Outcomes

Our subgroup analysis did not evidence variables to significantly impact the performance of the models. Nonetheless, when looking at the most performing classifiers, we observed a better trend for non-linear algorithms such as SVM and CNN, as shown in Fig. 4.

Fig. 4.

SROC curve for the two classification tasks considered together. From this graph, we can assess that the non-linear classifiers tend to perform better with respect to simpler solutions such as logistic regression classifiers. This can be evinced from the position of the non-linear classifiers (CNN, SVM, RF, and CNN + SVM) which are more towards the top-left corner of the graph, with respect to the linear (LR).

We furthermore investigated the following variables, which may significantly impact the model performance: (i) the selected features, which are the final data representation, in a numerical form, fed to the AI algorithm to perform the given classification task; (ii) the MRI modalities from which the input images are obtained. Regarding features, the most used were the textural (28.6%), followed by the deep (20.4%) and voxel intensities features (12.2%). Fig. 5A provides a visual representation of the features distribution among the studies included in our review. In this figure, each square corresponds to a single radiomic model. Note that in a single study, more than one radiomic model may be described, each one tailoring a different and specific classification task. Concerning the employed MRI modality, the CE-T1W was the most reported MRI sequence (36.5%), followed by T2-FLAIR (15.6%), and DWI (14.6%). Fig. 5B shows the distribution of the MRI modalities among the included studies. A majority of studies (58.1%) used a cross-validated analysis.

Fig. 5.

Radiomics features. (A) A visual representation of the radiomics features distribution among the studies included in our review. Each square corresponds to the features fed as input to a single radiomic model. Note that in a single study, more than one radiomic model may be described, each one tailoring a different and specific classification task. A final remark must be made concerning the so-called “deep features from Imagenet”. This is indeed a widely used data representation in the context of deep learning, as detailed in the corresponding studies. (B) Pie chart showing the distribution of the MRI modalities among the included studies. Note that a single radiomic model may be fed with data obtained from different MRI modalities.

3.5 Study Heterogeneity

The χ2 test suggested substantial heterogeneity of SEN and SPE, for both LGGs vs. HGGs and WHO III vs. IV categorizations.

4. Discussion

We provided an overview of the performance of the current radiomics models for glioma grading prediction. The overall performance resulted higher for the HGGs vs. LGGs discrimination task than the WHO grade 3 vs. 4 task, both in terms of SPE and SEN. The FPR was higher than the false negative rate (FNR) for both differentiation tasks indicating a greater capability to rule out HGGs or WHO grade 4 gliomas rather than identifying these entities. The studied variables did not impact significantly the performance, but we observed a better trend for non-linear classifiers such as SVM and CNN.

4.1 Radiomics Models

As shown in Fig. 6, there has been an outstanding growth of expertise in AI developments and laboratories, which has brought on a significantly higher number of studies published over the years.

Fig. 6.

Cumulative number of radiomics studies included in our study over years.

Radiomics has captured the interest of neuroradiologists and neuro-oncologists, who progressively upheld the research field [12, 57]. On the one hand, novel strategies for feature extraction and segmentation methods have been developed that reverberated the impact of radiomics on many fields of neuro-oncology and neuroradiology, multiplying the possible clinical applications [4, 7, 9, 11]. On the other hand, the concept of DL, as a method of data representational learning able to learn from end to end by itself without requiring any handcrafted features or human-based data representation, originated as an innovative branch of radiomics [58]. Even though a minority of the articles focused on DL, many hybrid studies were included that imbricate the ML classifiers with novel DL algorithms, accomplishing fully automatic detection processes. All studies included in our meta-analysis extracted features (handcrafted or DL-based) from MRI sequences, but other imaging modalities have been used for different purposes, particularly positron emission tomography in radiogenomics [59, 60].

4.2 Radiomics Performance

Our data confirm the paramount role of radiomics in predicting the grade of cerebral gliomas, as suggested by previous authors [61]. The overall performance reflects the great predictivity of the current models included in the meta-analysis.

We found a higher prediction capability of radiomics for the HGGs vs. LGGs differentiation task than the WHO grade 3 vs. grade 4 task. In accordance with the current literature, the main prognostic impact of glioma grading is provided by the LGGs vs. HGGs differentiation [62, 63, 64]. Even though the definitive diagnosis of gliomas is histopathological, these data strongly support the potential role of radiomics for an initial diagnostic orientation in the definition of grade glioma. This may be particularly important for those patients with doubtful neuroradiological imaging where there is no clear orientation toward a diagnosis of LGG vs. HGG [65, 66, 67, 68].

When looking at the SROC curves, we observed a smaller and better-performing prediction region in the LGGs vs. HGGs differentiation task, while the prediction region of grade 3 vs. grade 4 resulted to be larger and more oriented toward superior FPRs. This suggests that although we have not witnessed any advancements in performance thus far, we have reason to believe that we may observe progress in the upcoming years in distinguishing between LGGs and HGGs. This is shown by the SROC curve, which illustrates the potential for future improvement.

4.3 Determinants of Performance

Given the blossoming of AI models over the last decade, we expected the year of publication to significantly impact the performance of the models. Yet, we did not find a linear correlation between time of development and performance.

Textural features, along with deep and voxel intensities features were the most performing. Feature extraction is driven by algorithms to select ones appropriate for a precise task [9]. For glioma grading, textural features were reported as the best-performing. These common features, quantify the spatial variation of grey-level intensity inferring image heterogeneity [17, 69]. Frequent textural features are energy, entropy, inertia, correlation, and others [47, 50]. Zhou et al. [56] identified gray-level co-occurrence matrix (GLCM)-homogeneity as the main texture feature to predict histological grade from CE-T1W images. Likewise, Liu et al. [37] found two GLCM texture features to reflect the glioma grade. The multiparametric texture analysis of Hashido et al. [31] also included the GLCM features. Their study found that GLCM-based entropy effectively differentiated between LGGs and HGGs in PWI. Skogen et al. [17] used MRI texture analysis (MRTA) to assess tumor heterogeneity. Extraction of texture features at fine anatomical scales best discriminated LGG and HGG.

The most reported best-performing MRI sequences were the CE-T1W, T2-FLAIR, and DWI. The CE-T1W and T2-FLAIR modalities were particularly regarded for the HGGs vs. LGGs differentiation task, while the DWI has mostly been reported as the best-performing sequence for the WHO grade 3 vs. grade 4 discrimination task. Apparent diffusion coefficient (ADC) maps calculated from DWI, along with dynamic contrast-enhanced (DCE), dynamic susceptibility contrast (DSC), and arterial spin labeling (ASL) PWI have been increasingly reported as promising alternatives and will perhaps be the sequences most chosen for feature extraction over the next years [27, 30, 43, 45].

Future models should not disregard external validation to ascertain a sufficient level of reproducibility and reliability. The lack of standardization and the customizability of radiomics reduce the applicability of AI in daily practice. Clear routes of development should be delineated to overcome the diversities in radiomics laboratories. Hopefully, this meta-analysis will orient forthcoming models to more defined and shared processes that will be possibly implemented in clinical practice.

4.4 The Role of Radiomics in the Prediction of Cellular and Molecular Patterns of Gliomas

Gliomas, a collection of primary brain tumors, result from the abnormal proliferation of glial cells, including ependymal cells, oligodendrocytes, and astrocytes. These tumors display distinct cellular characteristics based on their histological subtype. For example, astrocytomas, which make up a significant portion of gliomas, primarily consist of proliferating astrocytes. Oligodendrogliomas, on the other hand, stem from oligodendrocyte precursor cells and are identified by their “fried-egg” appearance, characterized by round nuclei and clear cytoplasm. Ependymomas, the third major subtype, develop from ependymal cells that line the brain and spinal cord ventricles. Each glioma subtype exhibits unique cellular features crucial for precise diagnosis and classification.

The classification and comprehension of gliomas heavily rely on their cellular characteristics. However, the 2016 WHO classification ushered in a transformative era of understanding these tumors through their molecular traits, carrying substantial implications for diagnosis, prognosis, and therapy selection [70]. A pivotal molecular anomaly identified in gliomas is the isocitrate dehydrogenase (IDH) mutation, especially the IDH1 and IDH2 mutations, which manifest in a subset of gliomas. These mutations are linked to distinct clinical and histological attributes and wield a critical role in glioma classification and management. Another crucial molecular alteration in glioma involves the loss of the tumor suppressor gene TP53, contributing to the pathogenesis of high-grade gliomas. Furthermore, the methylation status of the O-6-methylguanine-DNA methyltransferase (MGMT) gene promoter stands as a noteworthy molecular marker that impacts the response to alkylating chemotherapy agents. Molecular profiling has empowered the refinement of glioma categorization into more precise groups, providing guidance for treatment decisions and enhancing prognostic accuracy [71, 72, 73].

The 2021 WHO classification of gliomas brought significant innovations to the molecular classification of these brain tumors, offering a more detailed and comprehensive approach. It emphasizes the importance of integrating both histological and molecular data to establish an integrated diagnosis, leading to a more precise glioma classification. These molecular markers play a pivotal role in refining glioma classification, offering valuable insights for prognosis, treatment planning, and personalized therapeutic strategies. The inclusion of telomerase reverse transcriptase (TERT) promoter mutation, H3K27M mutation, v-raf murine sarcoma viral oncogene homolog B1 (BRAF)-fusion mutation, and the increased focus on O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation reflects the evolving understanding of glioma biology and the growing need for more accurate diagnosis and management [2].

In this context, radiomics can assume a central role in unveiling the cellular and molecular patterns of gliomas. Utilizing advanced radiomic analyses, which are based on quantitative characteristics derived from medical images, offers a non-invasive approach to assessing the tumor’s heterogeneity, microenvironment, and genetic attributes. For instance, research such as Kickingereder et al. [74] demonstrates that radiomic features can capture variations in cell density, vascularity, and necrosis within gliomas, effectively reflecting their histological and cellular diversity. Furthermore, radiomics can aid in the identification of crucial genetic and molecular markers, like IDH mutations and 1p/19q co-deletion [75]. Additionally, radiomic features may also mirror the molecular diversity of gliomas, contributing to more precise diagnoses, treatment planning, and patient stratification, as highlighted by Lambin et al. [6]. However, it is worth noting that at present, there is still a scarcity of radiomic studies that indicate predictive features for the molecular sub-classifications proposed in the WHO 2021 classification.

4.5 Limitations

The meta-analysis limiting in nature since it was mostly based on retrospective cohort studies, even if the number of patients was considerable. Due to the limited data available, we could not ascertain the performance of radiomics in other discrimination tasks, such as WHO grade 1 vs. grade 2, and grade 3 vs. grade 4. Given the bivariate model of the meta-analysis, we did not calculate the overall accuracy for the differentiation tasks. Nonetheless, to the best of our knowledge, this is the first meta-analysis to picture the current performance of radiomics for glioma grading, providing cutting-edge conclusions to pilot future models.

Moreover, a limitation possibly affecting the generalizability of radiomics studies for classification tasks consists of the data drift phenomenon. Data drift in radiomics can occur due to changes over time in the classification used in radiomics analysis, in this specific case represented by the CNS WHO Classification of Tumors [1]. This can lead to a loss of accuracy and reliability in radiomics models trained on the old classification criteria, which may no longer apply to the new classification [76]. It is worth noting that when it comes to radiomics models, the classification criteria used can also have an impact on their effectiveness. A recent study by Moodi et al. [77] found that ML algorithms delivered superior results in grading gliomas based on WHO 2021 criteria, as compared to the WHO 2016 classification criteria. To mitigate this type of data drift, it is important to update radiomics models and retrain them on the new classification criteria. It is also important to carefully track any changes in classification criteria and ensure that they are well documented so that radiomics analysis can be properly adjusted and validated accordingly [78].

5. Conclusions

This meta-analysis suggests that the current radiomics models perform better in distinguishing between LGGs and HGGs than between WHO grade 3 and WHO grade 4 gliomas, in terms of both SPE and SEN. Enhanced future models with increased accuracy can prove to be of clinical use for categorizing HGGs and LGGs.

MRI is the most preferred imaging method, and the CE-T1W sequence is found to be the most effective for current radiomics models. Textural features are commonly used, and modern non-linear classifiers show a promising trend.

Abbreviations

1H-MRS, proton magnetic resonance spectroscopy; ADC, apparent diffusion coefficient; AI, artificial intelligence; ASL, arterial spin labeling; CE-T1W, contrast-enhanced T1W; CIs, confidence intervals; CNN, convolutional neural network; CNS, central nervous system; DCE, dynamic contrast-enhanced; DL, deep learning; DSC, dynamic susceptibility contrast; DWI, diffusion-weighted imaging; ENR, elastic net regression; FN, false negative; FNR, false negative rate; FP, false positive; FPR, false positive rate; GAN, generative adversarial network; GB, gradient boost; GLCM, gray-level co-occurrence matrix; HGG, high-grade glioma; IDH, isocitrate dehydrogenase; LDA, linear discriminant analysis; LGG, low-grade glioma; LR, logistic regression; MGMT, O-6-methylguanine-DNA methyltransferase; ML, machine learning; MLP, multilayer perceptron; MRTA, MRI texture analysis; NB, naïve bayes; NN, nearest neighbors; NOS, Newcastle-Ottawa scale; PRISMA, preferred reporting items for systematic reviews and meta-analysis; PWI, perfusion-weighted imaging; RF, random forest; SE N, sensitivity; SPE, specificity; SROC, summary receiver operating characteristic; SVM, support vector machine; T2-FLAIR, T2W-fluid-attenuated inversion recover; TN, true negative; TP, true positive; WHO, world health organization.

Availability of Data and Materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Author Contributions

Conceptualization, EA, MZ, LDM, MMF and PPP; methodology, EA, MZ, LDM, FP, HC, KS, IT, MG, LU and WB; validation, LDM, EA, FP, LU, MG and MMF; formal analysis, LDM, EA, PPP and FP; investigation, LDM, EA, PPP and FP; resources, EA, MZ, LDM, FP, HC, KS, IT, MG, PPP and LU; data curation, LDM, EA, PPP and FP; writing—original draft preparation, LDM, EA, PPP and FP; writing—review and editing, EA, LDM, FP, MG, MMF, WB and LU; visualization, EA, LDM, FP, MG, MMF, WB, TI, MZ, PPP and LU; supervision, EA, LDM, PPP, MMF and WB; project administration, LDM, EA, PPP and MMF. All authors contributed to editorial changes in the manuscript. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

Ethics Approval and Consent to Participate

Not applicable.

Acknowledgment

Not applicable.

Funding

This research received no external funding.

Conflict of Interest

The authors declare no conflict of interest.

References
[1]
WHO Classification of Tumours Editorial Board. Central Nervous System Tumours. 5th edn. World Health Organization: Geneva, Switzerland. 2022.
[2]
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-oncology. 2021; 23: 1231–1251.
[3]
Gritsch S, Batchelor TT, Gonzalez Castro LN. Diagnostic, therapeutic, and prognostic implications of the 2021 World Health Or-ganization classification of tumors of the central nervous system. Cancer. 2022; 128: 47–58.
[4]
Abdel Razek AAK, Alksas A, Shehata M, AbdelKhalek A, Abdel Baky K, El-Baz A, et al. Clinical applications of artificial intelligence and radiomics in neuro-oncology imaging. Insights into Imaging. 2021; 12: 152.
[5]
Nandu H, Wen PY, Huang RY. Imaging in neuro-oncology. Therapeutic Advances in Neurological Disorders. 2018; 11: 1756286418759865.
[6]
Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RGPM, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. European Journal of Cancer (Oxford, England: 1990). 2012; 48: 441–446.
[7]
Yi Z, Long L, Zeng Y, Liu Z. Current Advances and Challenges in Radiomics of Brain Tumors. Frontiers in Oncology. 2021; 11: 732196.
[8]
Cuocolo R, Ugga L, Solari D, Corvino S, D’Amico A, Russo D, et al. Prediction of pituitary adenoma surgical consistency: radiomic data mining and machine learning on T2-weighted MRI. Neuroradiology. 2020; 62: 1649–1656.
[9]
Aboian M, Bousabarah K, Kazarian E, Zeevi T, Holler W, Merkaj S, et al. Clinical implementation of artificial intelligence in neuroradiology with development of a novel workflow-efficient picture archiving and communication system-based automated brain tumor segmentation and radiomic feature extraction. Frontiers in Neuroscience. 2022; 16: 860208.
[10]
Ugga L, Perillo T, Cuocolo R, Stanzione A, Romeo V, Green R, et al. Meningioma MRI radiomics and machine learning: systematic review, quality score assessment, and meta-analysis. Neuroradiology. 2021; 63: 1293–1304.
[11]
Xu J, Meng Y, Qiu K, Topatana W, Li S, Wei C, et al. Applications of Artificial Intelligence Based on Medical Imaging in Glioma: Current State and Future Challenges. Frontiers in Oncology. 2022; 12: 892056.
[12]
Park JE, Kim HS, Kim D, Park SY, Kim JY, Cho SJ, et al. A systematic review reporting quality of radiomics research in neuro-oncology: toward clinical utility and quality improvement using high-dimensional imaging features. BMC Cancer. 2020; 20: 29.
[13]
Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ (Clinical Research Ed.). 2021; 372: n71.
[14]
Ottawa Hospital Research Institute. Available at: http://www.ohri.ca/programs/clinical_epidemiology/oxford.asp (Accessed: 12 April 2022).
[15]
Luchini C, Stubbs B, Solmi M, Veronese N. Assessing the quality of studies in meta-analyses: Advantages and limitations of the Newcastle Ottawa Scale. World Journal of Meta-Analysis. 2017; 5: 80–84.
[16]
Doebler P, Holling H. Meta-analysis of Diagnostic Accuracy and ROC Curves with Covariate Adjusted Semiparametric Mixtures. Psychometrika. 2015; 80: 1084–1104.
[17]
Skogen K, Schulz A, Dormagen JB, Ganeshan B, Helseth E, Server A. Diagnostic performance of texture analysis on MRI in grading cerebral gliomas. European Journal of Radiology. 2016; 85: 824–829.
[18]
Vamvakas A, Williams SC, Theodorou K, Kapsalaki E, Fountas K, Kappas C, et al. Imaging biomarker analysis of ad-vanced multiparametric MRI for glioma grading. Physica Medica: PM: an International Journal Devoted to the Applications of Physics to Medicine and Biology: Official Journal of the Italian Association of Biomedical Physics (AIFB). 2019; 60: 188–198.
[19]
Cho HH, Lee SH, Kim J, Park H. Classification of the glioma grading using radiomics analysis. PeerJ. 2018; 6: e5982.
[20]
Chen W, Liu B, Peng S, Sun J, Qiao X. Computer-Aided Grading of Gliomas Combining Automatic Segmentation and Radi-omics. International Journal of Biomedical Imaging. 2018; 2018: 2512037.
[21]
Cheng J, Gao M, Liu J, Yue H, Kuang H, Liu J, et al. Multimodal Disentangled Variational Autoencoder With Game Theoretic Interpretability for Glioma Grading. IEEE Journal of Biomedical and Health Informatics. 2022; 26: 673–684.
[22]
Cheng J, Liu J, Yue H, Bai H, Pan Y, Wang J. Prediction of Glioma Grade Using Intratumoral and Peritumoral Radiomic Fea-tures From Multiparametric MRI Images. IEEE/ACM Transactions on Computational Biology and Bioinformatics. 2022; 19: 1084–1095.
[23]
Hwan-Ho Cho, Hyunjin Park. Classification of low-grade and high-grade glioma using multi-modal image radiomics features. Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference. 2017; 2017: 3081–3084.
[24]
Ding J, Zhao R, Qiu Q, Chen J, Duan J, Cao X, et al. Developing and validating a deep learning and radiomic model for glioma grading using multiplanar reconstructed magnetic resonance contrast-enhanced T1-weighted imaging: a robust, multi-institutional study. Quantitative Imaging in Medicine and Surgery. 2022; 12: 1517–1528.
[25]
Ditmer A, Zhang B, Shujaat T, Pavlina A, Luibrand N, Gaskill-Shipley M, et al. Diagnostic accuracy of MRI texture analysis for grading gliomas. Journal of Neuro-oncology. 2018; 140: 583–589.
[26]
Gao M, Huang S, Pan X, Liao X, Yang R, Liu J. Machine Learning-Based Radiomics Predicting Tumor Grades and Expression of Multiple Pathologic Biomarkers in Gliomas. Frontiers in Oncology. 2020; 10: 1676.
[27]
Gihr GA, Horvath-Rizea D, Hekeler E, Ganslandt O, Henkes H, Hoffmann KT, et al. Histogram Analysis of Diffusion Weighted Imaging in Low-Grade Gliomas: in vivo Characterization of Tumor Architecture and Corresponding Neuropathology. Frontiers in Oncology. 2020; 10: 206.
[28]
Guo J, Ren J, Shen J, Cheng R, He Y. Do the combination of multiparametric MRI-based radiomics and selected blood inflam-matory markers predict the grade and proliferation in glioma patients? Diagnostic and Interventional Radiology (Ankara, Turkey). 2021; 27: 440–449.
[29]
Gutta S, Acharya J, Shiroishi MS, Hwang D, Nayak KS. Improved Glioma Grading Using Deep Convolutional Neural Networks. AJNR. American Journal of Neuroradiology. 2021; 42: 233–239.
[30]
Hashido T, Saito S, Ishida T. Radiomics-Based Machine Learning Classification for Glioma Grading Using Diffusion- and Per-fusion-Weighted Magnetic Resonance Imaging. Journal of Computer Assisted Tomography. 2021; 45: 606–613.
[31]
Hashido T, Saito S, Ishida T. A radiomics-based comparative study on arterial spin labeling and dynamic susceptibility contrast perfusion-weighted imaging in gliomas. Scientific Reports. 2020; 10: 6121.
[32]
Hu Z, Zhuang Q, Xiao Y, Wu G, Shi Z, Chen L, et al. MIL normalization – prerequisites for accurate MRI radiomics analysis. Computers in Biology and Medicine. 2021; 133: 104403.
[33]
Huang WY, Wen LH, Wu G, Hu MZ, Zhang CC, Chen F, et al. Comparison of Radiomics Analyses Based on Different Magnetic Resonance Imaging Sequences in Grading and Molecular Genomic Typing of Glioma. Journal of Computer Assisted Tomogra-phy. 2021; 45: 110–120.
[34]
Kobayashi K, Miyake M, Takahashi M, Hamamoto R. Observing deep radiomics for the classification of glioma grades. Scientific Reports. 2021; 11: 10942.
[35]
Li Y, Ammari S, Lawrance L, Quillent A, Assi T, Lassau N, et al. Radiomics-Based Method for Predicting the Glioma Subtype as Defined by Tumor Grade, IDH Mutation, and 1p/19q Codeletion. Cancers. 2022; 14: 1778.
[36]
Lin K, Cidan W, Qi Y, Wang X. Glioma grading prediction using multiparametric magnetic resonance imaging-based radiomics combined with proton magnetic resonance spectroscopy and diffusion tensor imaging. Medical Physics. 2022; 49: 4419–4429.
[37]
Liu X, Li J, Liao X, Luo Z, Xu Q, Pan H, et al. Radiomics-based MRI for predicting Erythropoietin-producing hepato-cellular receptor A2 expression and tumor grade in brain diffuse gliomas. Neuroradiology. 2022; 64: 323–331.
[38]
Lu CF, Hsu FT, Hsieh KLC, Kao YCJ, Cheng SJ, Hsu JBK, et al. Machine Learning-Based Radiomics for Molecular Subtyping of Gliomas. Clinical Cancer Research: an Official Journal of the American Association for Cancer Research. 2018; 24: 4429–4436.
[39]
Nakamoto T, Takahashi W, Haga A, Takahashi S, Kiryu S, Nawa K, et al. Publisher Correction: Prediction of malignant glioma grades using contrast-enhanced T1-weighted and T2-weighted magnetic resonance images based on a radiomic analysis. Scientific Reports. 2020; 10: 3073.
[40]
Ning Z, Luo J, Xiao Q, Cai L, Chen Y, Yu X, et al. Multi-modal magnetic resonance imaging-based grading analysis for gliomas by integrating radiomics and deep features. Annals of Translational Medicine. 2021; 9: 298.
[41]
Park YW, Choi YS, Ahn SS, Chang JH, Kim SH, Lee SK. Radiomics MRI Phenotyping with Machine Learning to Predict the Grade of Lower-Grade Gliomas: A Study Focused on Nonenhancing Tumors. Korean Journal of Radiology. 2019; 20: 1381–1389.
[42]
Reza SMS, Samad MD, Shboul ZA, Jones KA, Iftekharuddin KM. Glioma grading using structural magnetic resonance imaging and molecular data. Journal of Medical Imaging (Bellingham, Wash.). 2019; 6: 024501.
[43]
Su C, Chen X, Liu C, Li S, Jiang J, Qin Y, et al. T2-FLAIR, DWI and DKI radiomics satisfactorily predicts histological grade and Ki-67 proliferation index in gliomas. American Journal of Translational Research. 2021; 13: 9182–9194.
[44]
Su C, Jiang J, Zhang S, Shi J, Xu K, Shen N, et al. Radiomics based on multicontrast MRI can precisely differentiate among glioma subtypes and predict tumour-proliferative behaviour. European Radiology. 2019; 29: 1986–1996.
[45]
Sudre CH, Panovska-Griffiths J, Sanverdi E, Brandner S, Katsaros VK, Stranjalis G, et al. Machine learning assisted DSC-MRI radiomics as a tool for glioma classification by grade and mutation status. BMC Medical Informatics and Decision Making. 2020; 20: 149.
[46]
Takahashi S, Takahashi W, Tanaka S, Haga A, Nakamoto T, Suzuki Y, et al. Radiomics Analysis for Glioma Malignancy Evaluation Using Diffusion Kurtosis and Tensor Imaging. International Journal of Radiation Oncology, Biology, Physics. 2019; 105: 784–791.
[47]
Tian Q, Yan LF, Zhang X, Zhang X, Hu YC, Han Y, et al. Radiomics strategy for glioma grading using texture features from multiparametric MRI. Journal of Magnetic Resonance Imaging: JMRI. 2018; 48: 1518–1528.
[48]
van der Voort SR, Incekara F, Wijnenga MMJ, Kapsas G, Gahrmann R, Schouten JW, et al. Combined molecular sub-typing, grading, and segmentation of glioma using multi-task deep learning. Neuro-oncology. 2023; 25: 279–289.
[49]
Wang Q, Li Q, Mi R, Ye H, Zhang H, Chen B, et al. Radiomics Nomogram Building From Multiparametric MRI to Pre-dict Grade in Patients With Glioma: A Cohort Study. Journal of Magnetic Resonance Imaging: JMRI. 2019; 49: 825–833.
[50]
Xie T, Chen X, Fang J, Kang H, Xue W, Tong H, et al. Textural features of dynamic contrast-enhanced MRI derived model-free and model-based parameter maps in glioma grading. Journal of Magnetic Resonance Imaging: JMRI. 2018; 47: 1099–1111.
[51]
Xu C, Peng Y, Zhu W, Chen Z, Li J, Tan W, et al. An automated approach for predicting glioma grade and survival of LGG patients using CNN and radiomics. Frontiers in Oncology. 2022; 12: 969907.
[52]
Xu J, Ren Y, Zhao X, Wang X, Yu X, Yao Z, et al. Incorporating multiple magnetic resonance diffusion models to dif-ferentiate low- and high-grade adult gliomas: a machine learning approach. Quantitative Imaging in Medicine and Surgery. 2022; 12: 5171–5183.
[53]
Zhang Z, Xiao J, Wu S, Lv F, Gong J, Jiang L, et al. Deep Convolutional Radiomic Features on Diffusion Tensor Images for Classification of Glioma Grades. Journal of Digital Imaging. 2020; 33: 826–837.
[54]
Zhao SS, Feng XL, Hu YC, Han Y, Tian Q, Sun YZ, et al. Better efficacy in differentiating WHO grade II from III oli-godendrogliomas with machine-learning than radiologist’s reading from conventional T1 contrast-enhanced and fluid attenuated inversion recovery images. BMC Neurology. 2020; 20: 48.
[55]
Zhou H, Xu R, Mei H, Zhang L, Yu Q, Liu R, et al. Application of Enhanced T1WI of MRI Radiomics in Glioma Grad-ing. International Journal of Clinical Practice. 2022; 2022: 3252574.
[56]
Zhou H, Vallières M, Bai HX, Su C, Tang H, Oldridge D, et al. MRI features predict survival and molecular markers in diffuse lower-grade gliomas. Neuro-oncology. 2017; 19: 862–870.
[57]
Rudie JD, Rauschecker AM, Bryan RN, Davatzikos C, Mohan S. Emerging Applications of Artificial Intelligence in Neu-ro-Oncology. Radiology. 2019; 290: 607–618.
[58]
Zhang D, Zhu W, Guo J, Chen W, Gu X. Application of artificial intelligence in glioma researches: A bibliometric analysis. Fron-tiers in Oncology. 2022; 12: 978427.
[59]
Ahrari S, Zaragori T, Bros M, Oster J, Imbert L, Verger A. Implementing the Point Spread Function Deconvolution for Better Molecular Characterization of Newly Diagnosed Gliomas: A Dynamic 18F-FDOPA PET Radiomics Study. Cancers. 2022; 14: 5765.
[60]
Zhang L, Pan H, Liu Z, Gao J, Xu X, Wang L, et al. Multicenter clinical radiomics-integrated model based on [18F]FDG PET and multi-modal MRI predict ATRX mutation status in IDH-mutant lower-grade gliomas. European Radiology. 2023; 33: 872–883.
[61]
Sohn CK, Bisdas S. Diagnostic Accuracy of Machine Learning-Based Radiomics in Grading Gliomas: Systematic Review and Meta-Analysis. Contrast Media & Molecular Imaging. 2020; 2020: 2127062.
[62]
Sharma A, Graber JJ. Overview of prognostic factors in adult gliomas. Annals of Palliative Medicine. 2021; 10: 863–874.
[63]
Liang J, Lv X, Lu C, Ye X, Chen X, Fu J, et al. Prognostic factors of patients with Gliomas - an analysis on 335 patients with Glioblastoma and other forms of Gliomas. BMC Cancer. 2020; 20: 35.
[64]
Marra JS, Mendes GP, Yoshinari GH, Jr, da Silva Guimarães F, Mazin SC, de Oliveira HF. Survival after radiation therapy for high-grade glioma. Reports of Practical Oncology and Radiotherapy: Journal of Greatpoland Cancer Center in Poznan and Polish Society of Radiation Oncology. 2019; 24: 35–40.
[65]
Molinaro AM, Hervey-Jumper S, Morshed RA, Young J, Han SJ, Chunduru P, et al. Association of Maximal Extent of Resection of Contrast-Enhanced and Non-Contrast-Enhanced Tumor With Survival Within Molecular Subgroups of Patients With Newly Diagnosed Glioblastoma. JAMA Oncology. 2020; 6: 495–503.
[66]
Chang SM, Parney IF, Huang W, Anderson FA, Jr, Asher AL, Bernstein M, et al. Patterns of care for adults with newly diagnosed malignant glioma. JAMA. 2005; 293: 557–564.
[67]
Scott JN, Brasher PMA, Sevick RJ, Rewcastle NB, Forsyth PA. How often are nonenhancing supratentorial gliomas malignant? A population study. Neurology. 2002; 59: 947–949.
[68]
Velázquez Vega JE, Brat DJ, Ryken TC, Olson JJ. The role of neuropathology in the management of newly diagnosed glioblas-toma: a systematic review and evidence-based clinical practice guideline. Journal of Neuro-oncology. 2020; 150: 143–164.
[69]
Haralick RM, Shanmugam K, Dinstein IH. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics. 1973; 610–621.
[70]
Louis DN, Perry A, Reifenberger G, von Deimling A, Figarella-Branger D, Cavenee WK, et al. The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary. Acta Neuropathologica. 2016; 131: 803–820.
[71]
Brennan CW, Verhaak RGW, McKenna A, Campos B, Noushmehr H, Salama SR, et al. The somatic genomic landscape of glioblastoma. Cell. 2013; 155: 462–477.
[72]
Wick W, Gorlia T, Bendszus M, Taphoorn M, Sahm F, Harting I, et al. Lomustine and Bevacizumab in Progressive Glio-blastoma. The New England Journal of Medicine. 2017; 377: 1954–1963.
[73]
Desjardins A, Gromeier M, Herndon JE, 2nd, Beaubier N, Bolognesi DP, Friedman AH, et al. Recurrent Glioblastoma Treated with Recombinant Poliovirus. The New England Journal of Medicine. 2018; 379: 150–161.
[74]
Kickingereder P, Burth S, Wick A, Götz M, Eidel O, Schlemmer HP, et al. Radiomic Profiling of Glioblastoma: Identify-ing an Imaging Predictor of Patient Survival with Improved Performance over Established Clinical and Radiologic Risk Models. Radiolo-gy. 2016; 280: 880–889.
[75]
Law M, Young RJ, Babb JS, Peccerelli N, Chheang S, Gruber ML, et al. Gliomas: predicting time to progression or sur-vival with cerebral blood volume measurements at dynamic susceptibility-weighted contrast-enhanced perfusion MR imaging. Radiology. 2008; 247: 490–498.
[76]
Hoens TR, Polikar R, Chawla NV. Learning from streaming data with concept drift and imbalance: an overview. Progress in Artificial Intelligence. 2012; 1: 89–101.
[77]
Moodi F, Khodadadi Shoushtari F, Ghadimi DJ, Valizadeh G, Khormali E, Salari HM, et al. Glioma Tumor Grading Us-ing Radiomics on Conventional MRI: A Comparative Study of WHO 2021 and WHO 2016 Classification of Central Nervous Tumors. Journal of Magnetic Resonance Imaging: JMRI. 2023; 10.1002/jmri.29146.
[78]
Parmar C, Grossmann P, Bussink J, Lambin P, Aerts HJWL. Machine Learning methods for Quantitative Radiomic Biomarkers. Scientific Reports. 2015; 5: 13087.

Publisher’s Note: IMR Press stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share
Back to top