Background
Pediatric low-grade gliomas (PLGGs) are the most common pediatric brain tumors, accounting for more than 30% of central nervous system (CNS) tumors in children [
1]. According to the 2016 World Health Organization (WHO) classification of CNS tumors, PLGGs comprise a histologically heterogenous group of Grade I and II tumors, including pilocytic astrocytoma (PA, Grade I), diffuse astrocytoma (AII, Grade II), oligodendroglioma (OII, Grade II), oligoastrocytoma (OAII, Grade II), pleomorphic xanthoastrocytoma (PXA, Grade II), dysembryoplastic neuroepithelial tumor (Grade I), neuronal-glial tumor (Grades I and II) and several others [
2]. In clinical practice, PLGGs are generally regarded as a single group of tumors with relatively quiescent biological behavior and favorable prognosis [
1‐
3]. Nevertheless, recurrence or progression still occurs in about 30% of PLGGs [
1,
2,
4]. Postoperative adjuvant therapies for PLGGs include radiation therapy and systematic chemotherapy, which may cause long-term morbidity and toxicity [
1].
Compared to adult low-grade gliomas, PLGGs have different features in molecular pathology [
2]. Most PLGGs possess alterations in RAS/MAPK pathway, in which
BRAF is a vital component [
2,
5,
6]. In our previous study using a large set of 289 PLGGs to investigate biomarkers of molecular pathology and their clinical significance, the
KIAA1549-BRAF fusion,
MYB amplification,
CDKN2A deletion,
BRAFV600E,
H3F3A,
TERT promoter mutations, and
ATRX loss were identified in PLGGs [
2]. Emphatically, the combination of the previous molecular markers has successfully categorized PLGGs into four molecular risk groups (low-risk, intermediate-I, intermediate-II, and high-risk) with distinct survivals [
2]. These findings highlight the importance of molecular stratification in evaluation and management of PLGGs.
Non-invasive prediction of molecular biomarkers or groups of gliomas is challenging [
7]. Recent progress on artificial intelligence (AI) algorithms has considerably promoted automatically quantifying radiologic patterns, and several clinically relevant molecular biomarkers or groups have been identified by leveraging on AI algorithms in adult gliomas [
8‐
11]. Recently, Wagner MW et al. developed and validated a radiomic signature that is predictive of the
BRAF status of PLGGs [
12]. However, there is a lack of study investigating the relationship between radiological features and risk groups of PLGGs defined by multiple molecular markers utilizing AI algorithms.
In the current study, radiomic features from multiparametric MRI, including T1-weighted, T1-weighted gadolinium contrast-enhanced, T2-weighted, fluid attenuated inversion recovery, and apparent diffusion coefficient images (T1, T1c, T2, FLAIR, and ADC), were extracted from 61 PLGG patients to construct models for prediction of molecular subgroups (low-risk group and intermediate/high-risk group) and molecular marker (KIAA1549-BRAF fusion) by leveraging machine learning algorithms. We aim to demonstrate that MRI patterns are significantly associated with key molecular biomarker and are able to predict molecular subgroups of PLGGs.
Discussion
As the most frequent brain tumors in children, PLGGs comprise a group of gliomas with heterogeneous histological types and different tumor locations [
16,
17]. In the past decades, novel biological insights into the genetic background of PLGGs have been acquired by extensive investigations [
5,
6,
18‐
21]. Unlike adult low-grade gliomas that are characterized by robust molecular alterations such as
IDH mutations, 1p/19q codeletion, and
TERT promoter mutations [
22‐
24], PLGGs harbor their own molecular alterations distinct from adult counterparts [
3]. It was revealed that nearly all PLGGs converge on the alterations of the MAPK pathway, with these alterations 100% existing in pediatric pilocytic astrocytoma [
5,
6]. The most common molecular alteration in the MAPK axis is
KIAA1549-BRAF fusion, which is caused by tandem duplication and rearrangement between
BRAF and
KIAA1549 at chromosome 7q34 [
2]. The
BRAF gene is also the most common point mutation target in PLGGs, the majority being V600E hotspot mutation [
19,
20]. Several studies have demonstrated that
KIAA1549-BRAF fusion predicts better survivals in patients with PLGGs [
2,
25], while
BRAFV600E point mutation is associated with inferior prognosis [
26]. Aside from predictive values,
BRAF gene alterations are also targets for novel drugs used in preliminary clinical trials for PLGGs.
BRAF inhibitor such as dabrafenib has shown a positive response rate in a multicenter phase I study including patients with PLGGs [
27]. Zhang J et al. reported 25% of diffuse cerebral gliomas in children carried abnormalities in
MYB and
MYBL1 using whole genome sequencing [
6]. Our previous study has identified
MYB amplification in 10.6% PLGGs and revealed this genetic alteration was associated with significantly longer survivals of PLGGs. Therefore, the presence of either
KIAA1549-BRAF fusion or
MYB amplification categorizes a low-risk subset of PLGGs with a favorable prognosis [
2].
Due to the limited number of intermediate (n = 23) and high-risk (n = 2) groups in the current set, we combined these two risk groups into one risk group (intermediate/high-risk group). Compared to the intermediate/high-risk group, the low-risk group of PLGGs confers an excellent survival with no mortality and rare tumor recurrence until the follow-up time point [
2]. Therefore, identifying this group of PLGGs is of considerable significance since it has the potential to aid clinical decision-making such as the selection of molecular targeted therapies or curtailment of postoperative adjuvant therapies. However, the assignment of PLGGs risk groups requires the detection of multiple molecular biomarkers (
BRAF fusion,
MYB amplification,
BRAFV600E mutation,
CDKN2A deletion,
TERTp mutation,
H3F3A mutation, and
ATRX loss), which are not all available in a vast number of medical centers with constrained resources. In addition, postoperative rather than the preoperative assignment of risk groups of PLGGs will inevitably lose the chance to guide personalized surgical resection strategies for these tumors. For instance, in low-risk group of PLGGs, it may be wise to perform palliative resection to preserve neurological function, rather than pursuing total tumor resection, since this group of PLGGs presents quiescent biological behavior.
Radiomics is an emerging research realm that investigates the relationship between radiographic features and tumor genotype, which serves as a promising approach to discriminate surrogate biomarkers with an accurate reflection of tumor genomics [
7]. Radiographic data from MRI of CNS tumors are also extensively investigated by radiomic strategies by leveraging AI algorithms, and adult gliomas are the most frequently studied CNS tumors [
7]. Specifically, machine learning or deep learning algorithms trained on preoperative MRI were demonstrated to predict molecular biomarkers such as
IDH mutations, 1p/19q codeletion, and
TERT promoter mutations, or molecular subgroups based on
IDH mutations and 1p/19q codeletion in adult gliomas with remarkable sensitivity and specificity [
8‐
11]. A previous study has revealed radiomics-based prediction of
BRAF status in PLGGs appears feasible [
12]. However, whether radiomic features could accurately reflect the risk group of PLGGs remains unexplored. Our results demonstrated preoperative MRI patterns were able to predict either molecular biomarker (
BRAF fusion) or risk group based on multiple molecular biomarkers, and yielded a satisfying performance, with AUC of 0.818 and 0.833, respectively.
It is worth noting that, identifying accurate and reproducible radiomic features of tumors is an essential step before translating into clinical application. As described in the previous literature, we employed a two-reader manual delineation approach by calculating the ICC between the same feature extracted from two VOIs to assess the feature reproducibility [
28‐
30]. With the advancement of computing power, the use of semi-automatic or automatic approach has also provided sufficiently reliable tumor segmentation and feature stability [
12,
31,
32].
To our knowledge, no prior study has investigated radiomic features of PLGGs using DWI. Our results indicate that the GLCM or GLDM texture features of ADC maps contribute to assess either molecular alteration or risk stratification in PLGGs. Likewise, it was reported that texture features and ADC parameters were important imaging markers to discriminate molecular subtypes in adult diffuse gliomas [
9,
33‐
35]. For instance, Kihira S et al. found that addition of GLCM texture features from diffusion images to conventional MRI features could improve the diagnostic performance in determination of
MGMT methylation status in gliomas [
35]. Meanwhile, in our previous study, the results showed that the GLCM texture features from ADC maps played an important role in predicting
IDH mutation and
TERT promoter mutation of gliomas, while GLDM texture features from ADC maps were important for 1p/19q codeletion [
9]. These phenomena may partly be explained by, that ADC values of brain tumors are inversely related to cellularity [
36], and that texture features quantify local image patterns reflecting subtle intratumoral heterogeneity [
33]. In addition, features from the conventional MR sequences were also revealed to play a role in the prediction model for
BRAF status or risk stratification of PLGGs. This may explain why in the internal validation set, we archived a higher AUC (mean AUC = 0.805) than the AUC (0.75) reported by the previous study using only FLAIR sequence for model development [
12]. It is reasonable to infer that a radiomic signature with features from multiparametric MRI is more effective and reliable than a single sequence.
Several limitations need to be pointed out in the current study. The first limitation is the relatively small sample size of the set, which hampers us to divide the intermediate-risk group from high-risk group for developing prediction model. Multi-institutional studies with larger sample size are necessary to further validate our findings. Second, advanced MR sequences such as diffusion tensor imaging (DTI), perfusion-weighted imaging (PWI), and diffusion kurtosis imaging (DKI) are welcome to further excavate the potential of MRI for predicting genotypes of PLGGs. Third, extensive integrative analysis on high through-put sequencing with paired MRI data, as well as in vivo imaging studies are required to clarify the elusive mechanisms on the relationship between radiomic patterns and genotypes of PLGGs. Lastly, manual tumor segmentation is a time-consuming and costly task. In future, we will employ semi-automated or automated segmentation algorithms to achieve accurate and repeatable tumor segmentation.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.