Glioma is a highly fatal disease that represents the most frequent form of primary cancer in the central nervous system (CNS), accounting for about 80% of all malignant tumors in the brain (1–3). Due to the same standardized treatment that can result in varying prognoses for different patients, it may be necessary to make specialized treatment decisions based on the tumor grade in clinical practice (4–6). Gliomas are categorized into grades I-IV in the Central Nervous System Midstream Classification of the World Health Organization (WHO) of 2021, with grades I-II being low-grade gliomas (LGG) and grades III-IV being high-grade gliomas (HGG) (7, 8). Accurate preoperative grading of gliomas is essential for assessing prognosis and developing individualized treatment plans, such as the extent of surgical resection, and the decision of postoperative chemoradiotherapy (9).

In contemporary clinical practice, gliomas are graded based on surgical or puncture histopathologic investigation (10). This diagnostic method is intrusive and slow, though. Furthermore, tissue biopsies from one area of the tumor might not be indicative of the histology of the entire tumor due to the recognized heterogeneity of gliomas and sampling error (11–13). High-precision noninvasive solutions that can offer preoperative grading information are therefore gaining popularity. Over the past decades, magnetic resonance imaging (MRI) has emerged as a crucial non-invasive diagnostic and assistant therapeutic technique for brain tumors, which is used to aid in differential diagnosis, guide treatment planning, and monitor therapy response (14–17). Nevertheless, competent radiologists may easily spot tumors from MRI sequences with the naked eye, gliomas are difficult to discriminate based on grade because of the variability and diversity of the tumors, which is undoubtedly a great challenge for imaging technology (18–20).

Radiomics, a burgeoning discipline, employing automated data mining algorithms to extract characteristics from medical images in a high-throughput manner, has been demonstrated notable advancements in the realm of medical imaging applications (21–24). These extracted features are then utilized by the machine to train itself and generate the anticipated desired output (25–27). Notably, recent progress has been achieved in the prediction of grading gliomas through the utilization of preoperative MRI scans and various machine learning methods. Gemini et al. evaluated the capacity of the Visually AcceSAble Rembrandt Images (VASARI) scoring system in predicting glioma grades and Isocitrate Dehydrogenase (IDH) status, with a possible application in machine learning (28). You et al. utilized traditional radiomics and the VASARI standard to construct a model determining glioma grade with an Area Under the Curve (AUC) of 0.966 (29). Wang et al. created and assessed a multiparametric MRI-based radiomics nomogram for predicting glioma grading (30). In recent years, deep learning has exhibited excellent performance with broader application prospects and deeper development in clinical applications. Voort et al. developed a single multi-task convolutional neural network that used preoperative MRI scans to predict the molecular subtype and grade of glioma, and the independent dataset evaluated that the approach achieved good performance and generalized well (31). Li et al. compared predictive models established by traditional radiomics and deep learning based on multiparametric MRI for grading gliomas and demonstrated that the latter performed better in most circumstances (32). While deep learning-based models have very respectable efficacy, it is commonly recognized as a black box that lacks satisfactory explanatory power. However, these machine learning approaches have focused mainly on the intratumoral region and neglected the role of the peritumoral environment in glioma grade.

The peritumoral environment holds great potential and may provide insightful information for clinical evaluation of the aggressive biological behavior of the tumor (33–35). Regarding intratumoral and peritumoral radiomic analysis, two main research approaches emerged. The first involved feature-fusion, where features from both intratumoral and peritumoral volumes of interest (VOIs) were separately extracted and then integrated. The second method was VOI-fusion, wherein the tumor region was expanded outward by a specific range to create a new VOI combining intra- and peri-tumoral areas. Radiomic features from this newly generated VOI were then extracted for subsequent analysis. For instance, Li et al. independently delineated the gross-tumor region and the peritumor region which was defined as the parenchyma that fell within a 2-cm distance to the tumor boundary (35). The radiomic features extracted from the two regions were merged and screened to build the radiomic model. Differently, Shi et al. expanded the originally segmented masks of VOIs by five radial distances outside the tumor at 1 mm intervals, creating five new VOIs (34). The findings indicated that the radiomic signature derived from peritumoral regions, specifically at dilated distances of 1 mm and 3 mm, demonstrated the most effective prediction performance in different MRI sequences, respectively. To date, although numerous studies have explored the radiomics of the peritumoral region, there has been a lack of a definitive study that has determined which among them is more persuasive and authoritative. The growth and infiltration of gliomas lead to the disruption of the blood-brain barrier. Consequently, there is a leakage of water, electrolytes, and proteins from peritumoral blood vessels, resulting in increased water content within the brain parenchyma, which contributes to the formation of peripheral edema (32, 36). Given that the tumor and the surrounding edematous area were closely interconnected, forming the microenvironment crucial for tumor cell growth and infiltration, it would be more reasonable to explore them as an integrated whole.

Drawing on the current state of research, the two methods were performed and compared in our study. We conducted an investigation focusing on the intratumor region and its surrounding peritumoral edema region of preoperative MRI scans to predict the grade of glioma. A total of 30 radiomic models, which depended on four different combinations of predictive features source (intratumoral VOI, peritumoral VOI, VOI-fusion, and feature-fusion), five different machine learning classifiers, and different MRI sequences, were established to select the optimal model, which was used for the construction of the nomogram for accurately predicting glioma grade, thereby assisting in the development of personalized treatment strategies for patients, ensuring they receive optimal benefits.

In this study, we developed variously dependable models to preoperatively predict grade glioma by using MRI images, which were constructed by radiomic features extracted from the intratumoral region and peritumoral edema region. The XGBoost machine learning classifier, incorporating features extracted from the combination of intratumoral and peritumoral VOIs of MRI, exhibited superior performance in distinguishing between LGG and HGG. The internal validation and independent external test underscored the robustness and generalizability of the model. Additionally, the radiomic models derived from VOI-fusion outperformed those derived from feature-fusion in our study, suggesting more extensive investigations into peritumoral radiomics were necessary to determine a more standardized research method and provide more theoretical support for radiomic studies. Finally, a nomogram based on the optimal machine learning model and clinical parameter was established to detect potential applications for predicting glioma grade in clinical practice.

MRI serves as a routine tool for preliminary diagnosing, treatment planning, and monitoring the treatment response of patients with glioma (39, 40). Recent studies have consistently demonstrated a robust association between radiomic features extracted from multiparametric MRI scans and various applications related to gliomas (2, 41). Kim et al. distinguished between glioblastoma and primary CNS lymphoma using multiparametric MRI sequences that included contrast-enhanced T1-weighted, T2-weighted, and diffusion-weighted imaging (42). Fifteen features were chosen from all imaging modalities, and the model rendered an AUC of 0.979, demonstrating the potent prediction potential of multi-parameter MRI. Nevertheless, the prediction of models based on a single MRI sequence was not performed and compared. In our study, we constructed and compared various radiomic models based on the intratumoral region of the single contrast-enhanced T1 sequence, the single T2 flair sequence, and the combination of two sequences. As expected, the radiomic models derived from dual-sequence MR imaging outperformed those solely based on single contrast-enhanced T1 or single T2 flair sequence. Hence, multicontrast MRI-based radiomics was poised to enhance the predictive capability for glioma grading compared to that of single MRI sequence.

Previous studies have demonstrated that the heterogeneity of gliomas extended beyond the tumor interior to include the peritumoral region, where approximately 90% of gliomas recurred (43, 44). Glioma cells interacted with molecules in the peritumoral area to cause hypoxia, angiogenesis, and tumor infiltration, which would ultimately accelerate the growth of gliomas (45). Consequently, the significant potential existed in the peritumoral environment, which could provide important information for evaluating the aggressive biological behavior of the tumor clinically (46, 47). In studies of other cancers, Ding et al. investigated the effect of peritumoral features for predicting sentinel lymph node metastasis in breast cancer (48). They created peritumoral regions by expanding tumor regions of interest at thicknesses of 2 mm, 4 mm, 6 mm, and 8 mm. By incorporating peritumoral features, the accuracy in the validation set increased from 0.704 to 0.796. Shan et al. developed a prediction model using peritumoral radiomic signatures extracted from a 2 cm peritumoral area, assessing its effectiveness in predicting the early recurrence of hepatocellular carcinoma post-curative treatment (49). In the validation cohort, ROC curves and decision curves revealed superior prediction efficiency and greater clinical benefits with the peritumoral model. Consequently, the extraction and integration of peritumoral and intratumoral features present a promising avenue. However, existing radiomic-based techniques for grading gliomas focused primarily on the interior of the tumor and less on the peritumoral environment.

Peripheral edema and peritumor in gliomas are two distinct concepts, the peritumor is typically an area within a specified radius around the tumor, whereas the peripheral edema around a glioma is irregular and frequently dispersed along the cerebral gyrus. Previous research has shown that the degree of peritumoral edema increases with the pathological grade and aggressiveness of glioma (14). It was demonstrated that individuals with severe edema (>10 mm) had mean or overall survival rates that were more than 50% lower than those with mild edema (8, 9, 44). Cheng et al. compared the predictive power of the peritumor and peripheral edema region for grading gliomas and found the most predictive features were extracted from the peritumor region within an immediate distance of 1 mm from the tumor core based on MRI scans (8). Notably, the study did not explore a predictive model based on the peripheral edema region in combination with the intratumoral region. Considering that the prognosis of gliomas was strongly correlated with the occurrence of peritumoral edema, we attempted to investigate the radiomic models based on peripheral edema for glioma grade.

Concerning the two main research approaches in intratumoral and peritumoral radiomic analysis, which was simplified as feature-fusion and VOI-fusion, definitive studies remained absent establishing a persuasive conclusion about which approach made more sense and produced more predictive features. Our study filled this gap by conducting a comparative analysis of the two methodologies for the first time. In the internal validation, the AUC of models based on both two methods showed no obvious differences, the difference in AUC was consistent across classifiers. In the external test, each categorical classifier on the VOI-fusion model outperformed that of the feature-fusion model respectively, which indicated that the model constructed by extracting features from the intratumoral and peritumoral regions as a whole yielded higher prediction efficiency. Regarding VOI fusion, the radiomic features, such as shape, first-order, and texture, extracted comprehensive information taking into account both intratumoral and peripheral edema region. Regarding feature fusion, individual peritumoral features exhibited minimal statistical variance in effectively distinguishing HGG from LGG, rendering them prone to elimination during screening. In our study, the constructed model just by feature fusion performed poorly in external tests, indicating unstable performance. This highlighted the need for further studies to unravel the intricacies of intratumoral and peritumoral radiomics.

We assessed 1288 radiomics features for every MRI sequence, and 7728 features in total, which was distinctly more than most recent findings and included all significant variables for radiomic analysis (30, 45). To identify the most optimal model for our dataset, we applied the six previously discussed categories of radiomic models across five classifiers: LR, SVM, XGBoost, DT, and MLP. This meticulous process ensured the exploration of fully optimized models best suited for our data. Meanwhile, previous studies have identified several clinical parameters crucial in distinguishing between LGG and HGG. Wang et al. selected clinical factors including age and sex as well as radiomic signature to develop a nomogram to predict glioma grading (30). It is well-recognized that HGG tends to be diagnosed in the elderly (31, 50). Consistently, we found that only the age parameter was statistically significant in predicting glioma grading, using the uni-multivariate analyses. Despite the lower predictive capacity of clinical model based solely on age feature compared to radiomic models, the nomogram amalgamating clinical parameter and radiomic models surpassed the predictive efficacy of either model in isolation. In terms of the predictive efficiency of various machine learning methods, Voort et al. utilized a deep learning model to predict the grade of glioma, achieving an AUC of 0.81 in the external testing cohort of 240 patients from 13 different institutes (31). Similarly, Li et al. distinguished LGG from HGG by developing deep convolutional neural network models, achieving an AUC of 0.89. In comparison, the nomogram in our study showed great performance with an AUC of 0.825 in the independent external testing cohort, which was equivalent to the state-of-the-art research aforementioned.

For all this, there were limitations in this study. First, our study required a larger sample size from more centers to make the findings more convincing. Second, only two MRI sequences were employed in this study. Some advanced parametric MRI scans, such as DWI and DTI, have shown powerful potential in tumor research, and new scanning techniques should be explored (7, 18, 36). Third, the VOIs in our study were manually annotated, which was time-consuming and laborious, and even prone to inaccurate annotation. Deep learning-based tumor segmentation methods are expected to be employed to improve the accuracy and reliability of image segmentation. Finally, our study lacked molecular subtyping of the samples, which was critical for the prognosis of gliomas, and planned to integrate such information in future studies.

In this work, we assessed the function of radiomic models of intratumoral and peripheral edema regions in MRI scans for predicting glioma grade and validated the methodology on an independent external test dataset, which provided a fresh viewpoint on the disease. The nomogram combined clinical parameter and the optimal radiomic model was efficient in glioma grade, and this non-invasive approach was expected to promote clinical research and guide the management of individualized glioma treatment.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by Tianjin Medical University General Hospital’s Institutional Ethics Committee. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because The requirement for written informed consent forms from all patients taking part in the study was waived because of the retrospective investigation.

RT: Data curation, Writing – original draft. CS: Methodology, Writing – original draft. CW: Writing – original draft. TZ: Conceptualization, Writing – review & editing.