Individual structural MRI abnormality indexes are associated with molecular, histological and cognitive indicators for frontal glioma patients

Purpose: The conventional examination of molecular characteristics for glioma requires surgical removal and could not be used in the surgery planning. The aim is to develop a method previewing glioma malignancy with regular structural MRI images. Methods: 52 glioma patients in frontal lobe (mean age 43.2±9.3 years, 34 male) and 117 healthy controls (mean age 32.6±9.8 years, 83 male) participated in the study. All patients underwent neurocognitive test and molecular examinations, including 1p/19q co-deletion, isocitrate dehydrogenase (IDH) mutation, telomerase reverse transcriptase (TERT) promoter mutation and O 6 -methylguanine-DNA methyltransferase (MGMT) promotor methylation. A series of individual structural abnormality indexes based on preoperative structural MRI were proposed and explored the associations with these clinical indicators. Results: Individual structural abnormality maps displayed that bilateral hippocampus, parahippocampus, insula, putamen and thalamus were constantly affected by glioma regardless of the histological grade, tumor hemisphere and molecular status. Higher grade glioma patients suffered more structural abnormalities, especially in the contralateral hemisphere and non-tumor regions. The molecular indicators, including IDH1 mutation, 1p/19q co-deletion, TERT promoter mutations and MGMT promoter methylation, as well as neurocognitive performance were all signicantly correlated to individual structural abnormality indexes. Conclusion: Our proposed individual structural abnormality indexes show great potentials to access the glioma pathological, neurocognitive and molecular indicators, which is very helpful for neurosurgeons to determine the personalized treatment strategies.


Introduction
Gliomas represent the majority of primary central nervous system (CNS) malignant tumors, and the surgical outcomes are dependent on the precise and comprehensive acquirement of various diagnostic information, which includes histological grades, molecular biomarkers and cognitive function. Since 2016, World Health Organization (WHO) updates the classi cation criteria for CNS tumors, and speci c molecular biomarkers are included for gliomas beyond histological characteristics [1]. For example, the incidence of isocitrate dehydrogenase (IDH) mutations, 1p/19q co-deletions, and telomerase reverse transcriptase (TERT) promoter mutations in glioma were predictors for better treatment outcome regardless of histological results [2][3][4]. In light of the crucial roles of theses molecular parameters in glioma management, examinations of IDH mutations,1p/19q co-deletions, and TERT promoter mutations have become the routine diagnostic items in many neuropathology centers. However, the conventional determination of these molecular characteristics requires surgical removal of tumor tissues. Therefore, a noninvasive method that could access IDH-mutants, 1p/19q-co-deletions and TERT promoter mutations before surgery would be more helpful in the treatment strategy selection and the prognostic prediction, especially for initially diagnosed patients and patients who do not prefer surgery as the treatment.
Medical imaging technologies such as magnetic resonance imaging (MRI) and computed tomography (CT) are routine ways to obtain macroscopic information about glioma for the advantages of low cost, low or no damage and convenience [5,6]. Besides, the microscopic characteristics concealed in medical images have also been discovered, such as Radiomics, which has been widely used to predict histological and molecular biomarkers for glioma patients [7][8][9]. However, these studies have several intrinsic disadvantages: 1. extract group-level but not individual-level (patient-speci c) features for classi cation or prediction; 2. only pay attention to the tumor regions, but the tumor related alterations in non-tumor regions are ignored. In this context, individualized imaging indexes that could delineate wholebrain alterations are highly desired for glioma patients before surgery.
Recently, Stoecklein et al. proposed a novel individual index based on the functional connectivity of resting state functional MRI (rs-fMRI), and found this individual index was positively correlated with WHO tumor grade and the IDH mutation status [10]. However, rs-fMRI is still not a clinical routine MRI scan, and its spatial resolution is commonly not very high. Moreover, rs-fMRI analysis is usually limited to gray matter (GM) but rarely within white matter (WM). In contrast, T1 sequence is a basic structural MRI scan, and plenty of analyzing tools have been developed for T1 images to extract morphological features such as cortical volume and thickness. If T1 could offer individualized diagnostic information about genotype, histological grade, and neurocognitive score, it will be largely helpful to personalized presurgical planning and prognosis prediction. Inspired by the previous study [11] in measuring individual structural abnormality, we extend to quantify both GM and WM abnormality at individual level, and propose series of individual structural abnormality indexes based on extracted individual GM/WM abnormality map. The associations between these individual structural indexes and histological, molecular, and neurocognitive indicators are explored respectively, and these indexes were additionally combined together through canonical correlation analysis to improve the associations.

Patients
The study was approved by the Institutional Review Board of Beijing Tiantan Hospital, Capital Medical University, Beijing, China (KY2020-048-01). The study was also registered at Chinese ClinicalTrial Registry (ChiCTR2000031805). All study procedures were in accordance with the Declaration of Helsinki. Written informed consent was obtained from all patients. Fifty-two glioma patients and one hundred seventeen healthy controls participated this study, which were listed in Table 1. All gliomas in the cohort were diagnosed according to the criteria of the WHO classi cation system in the revised version of 2016. Inclusion criteria were suspected, newly diagnosed glioma with only one cancer region in the brain and age over 18 years. Exclusion criteria were previous cranial surgery, neuropsychiatric comorbidities, and any contraindications to MR scanning such as metal implants. Neuropsychological testing was conducted prior to the rst MRI scan using the Montreal Cognitive Assessment (MOCA) test. Histological con rmation of the diagnosis was obtained by surgical resection. Molecular markers, including 1p/19q codeletion, IDH1/2 mutation, TERT promoter mutation and O 6 -methylguanine-DNA methyltransferase (MGMT) promotor methylation were all collected. Chromosomes 1 and 19 were analyzed by the uorescence in situ hybridization method, and the IDH1/2 mutation and TERT promoter mutation were detected by sequence analysis, both following a previously described protocol [12]. MGMT promoter methylation was assessed by methylation-speci c PCR as described previously by our team [13]. Patients were followed with routine clinical visits after initiation of therapy. All healthy controls were recruited from the local community and university students.

Structural MRI acquirement
All subjects were scanned with a Philips Ingenia 3.0T MRI scanner at Beijing Tiantan Hospital. For both glioma patients and healthy controls, T1 sequence was collected with the following parameters: TR: 6. Image processing The tumor region of every patient was extracted from T2-air images with two sequential steps: 1. launch an automatic segmentation by ITK-SNAP software; 2. manually correct the segmentation by experienced neurosurgeons. After the segmentation, individual T2-air image was co-registered to its corresponding T1 image by SPM software, and the transformation matrix was used on the segmented tumor to obtain the matched tumor region in T1 image space.
Individual T1 images were processed with CAT12 software to calculate the brain tissue volume. The skull stripping and correction for bias-eld inhomogeneities were subsequently conducted. After that, the whole brain was segmented into different tissue types, e.g. gray matter and white matter. Then, the segmented GM and WM images were normalized to the MNI standard space with a modulation manner by DARTEL algorithm. Finally, the normalized GM and WM images were smoothed with 4-mm FWHM Gaussian kernel.
Individual structural abnormality index Figure 1 illustrated the owchart of the proposed individual structural abnormality map. In order to calculate individual structural abnormality maps for each glioma patient, all healthy controls were rstly used to construct the normative brain volume model. In this model, general linear model (GLM) was adopted to discover the relationship between voxel-wise volume and variables including age, sex, and total intracranial volume (TIV) as the following equation.
Here, β 1 , β 2 , β 3 were weights for age, sex and TIV on the voxel-wise volume, and GM and WM volume models were respectively constructed.
Once the models had been created, the individual structural abnormality map for each glioma patient was calculated based on W score, which was calculated as following: After the calculation of W score for each patient, a cutoff threshold (|W|>3) was set to the individual W score map to generate the individual structural abnormality map. Notably, an individualized explicit mask was generated for every patient which combined the corresponding tissue (GM/WM) prior probability template (threshold: 0.2) with individual tumor mask. Through the individual structural abnormality map, we proposed a series of abnormality indexes (8 GM indexes and 8 WM indexes, 16 in total) to re ect the characteristics of structural damages induced by glioma, which were listed in detail in Table 2 The datasets generated and analysed during the current study are available from the corresponding author on reasonable request.

Results
Histology, molecular markers, and clinical course There were 52 glioma patients (mean age 43.2±9.3 years, 34 male) and 117 healthy controls (mean age 32.6±9.8 years, 83 male) that participated in the study. All patients were prospectively included between May 2019 and July 2020. The WHO histological grade, IDH1, TERT mutation, 1p/19q deletion, MGMT methylation were acquired for every patient, and Montreal Cognitive Assessment (MOCA) score was also recorded. Table 1 summarized the demographic and clinical information for all subjects.
Individual structural abnormality map could discover structural alterations not only within the tumor but also outside the tumor.
It is no doubt that glioma could induce whole brain alterations. Figure 2 illustrated several patients' T1 images and corresponding personalized GM/WM abnormality maps (mapping back into T1 space).
Obviously, the cerebral structural abnormalities not only lay in tumors but also in non-tumor regions, and all patients displayed both atrophies and enlargements in GM and WM. Although the patients varied in tumor sizes and tumor grades, the personalized GM abnormality maps could more precisely detect the tumor region than WM abnormality maps. In addition, high grade glioma patients were found with larger abnormalities in non-tumor regions. Figure 3 respectively displayed the overlapping regions for individual structural abnormality map, and it is clear that no matter the tumor is located at the left hemisphere or right hemisphere, several regions outside the tumor were consistently found abnormality, including bilateral hippocampus, insula, putamen, parahippocampus, thalamus.
Individual structural abnormality indexes were associated with clinical indicators, and combinations of these indexes would promote the correlation.

Discussion
In this study, we rstly depicted the structural abnormality characteristics for glioma patients through an individual level method, and proposed several quantitative indexes to re ect the individual abnormality severity, which were also found with high correlations with histological, molecular, and cognitive indicators. These ndings demonstrated their enormous potentials in providing multiview information for the preoperative assessment of glioma patients, which may be nally used for individualized precision treatment and prognosis evaluation for glioma patients.
The non-invasiveness and convenience of imaging technologies make them act as indispensable tools for presurgical assessment of glioma patients. Among them, contrast-enhanced MRI is currently served as the diagnostic golden standard for patients with suspected brain neoplasms, which is often supplemented with spectroscopy, PET, or diffusion imaging. In contrast, T1 image is a basic structural imaging sequence that has been widely collected in various populations. Specially, it could be easily gained for healthy controls while contrast-enhanced MRI or PET is usually not applicable for healthy controls. Therefore, these imaging modalities are usually used in a manner of visual inspection by experienced neurosurgeons. In addition, there are commonly two ways to analyze T1 images for glioma patients: 1. Use voxel/tensor/surface/deformation based morphometry methods to study the structural alterations, but these methods could be only conducted through statistical comparisons at the group level; 2. Use Radiomics methods to predict tumor phenotypes and clinical indicators, however, these methods only focus on the tumor region, hence couldn't provide information about non-tumor regions. Our method in the study solved the above-mentioned shortcomings, and further proposed a series of abnormality metrics at individual level to quantitatively delineate the structural alterations caused by glioma. Therefore, our method could offer comprehensive information about every glioma patient and may assist neurosurgeons to decide the most appropriate therapeutic strategy for every glioma patient.
Recently, Stoecklein et al. rstly demonstrated the feasibility of rs-fMRI to characterize individual glioma patient [10]. More importantly, this study manifested the tumor related changes in functional connectivity was unique for glioma patients with different histological grades, emphasizing the essentials of individual imaging biomarkers [10]. However, rs-fMRI is a type of technology with relatively low signal to noise ratio, and its test-retest reliability is also not high and dependent on the scanning durations, which makes it still challengeable for widespread applications for glioma patients. By contrast, structural MRI (sMRI) is with high image quality and test-retest reliability as well as appropriate scanning time. Besides, fMRI is usually limited to the gray matter while sMRI could detect the morphological changes in both gray matter and white matter. Moreover, our results discovered white matter indexes are as informative as gray matter indexes, and combining them together could improve the associations with clinical indicators of glioma patients. Finally, our study actually demonstrated the structural alterations in the brain were also unique among glioma patients. Taken together, glioma may lead to patient speci c alterations in both brain function and structure.
Although glioma patients display distinct structural abnormalities at the individual level, it is interesting to nd they also share common structural abnormalities in regions such as bilateral hippocampus, insula, putamen, thalamus and parahippocampus. Specially, these abnormal regions are not dependent on the hemisphere of tumor, histological grade and molecular indicators. On the one hand, it implies tumor related structural abnormalities are not only limited in the tumor hemisphere but also in the contralateral hemisphere no matter whether the tumor hemisphere is the dominant hemisphere. On the other hand, why these regions rather than other regions show abnormalities needs additional consideration. We speculate one possible reason is that all glioma patients in the study are with tumor in the frontal lobe, which participates in some key function networks (e.g. default mode network, attention network, salience network, affective network) together with these abnormal regions. Once the frontal lobe suffers from the glioma, these abnormal regions have to undergo some compensatory alterations to maintain these functions. Previous literatures have reported that as many as 90% of brain tumor patients would show tumor-related cognitive de cits (e.g., memory, attention, information processing, executive functioning impairments) [14,15], we infer this may be caused by the broken compensatory balance after the glioma surgery. At last, our ndings show that the combination of all structural indexes achieved a high correlation (r=-0.553) with MOCA score, indicating the cognitive ability is actually related to the structural abnormalities in glioma patients.
For the prognosis of glioma patients, the histological grade and IDH mutation are two vital known factors. For example, lower grade gliomas with wild-type IDH were reported to show similar prognosis with glioblastomas, and IDH mutated glioblastomas were found with better prognosis than IDH wild-type glioblastomas [16,17]. In addition, anaplastic gliomas with IDH wild-type have worse prognosis than glioblastomas with IDH mutation [16]. However, it is still di cult to timely know genetic examination results during the surgery, therefore the analysis of preoperative images has become the most possible way to predict these indicators. Several studies [18][19][20]7] using different modality images have been adopted for IDH prediction with high accuracy, but these models (e.g. deep learning) are hard to interpret intuitionally. Our individual structural abnormality indexes provide new perspective to access both histological grade and IDH mutation, and several single indexes show high correlations with histological grade and IDH mutation. When all these indexes are combined together, the correlational values are further improved (r=0.786 and 0.654 respectively). Our results indicate the potential of these individual structural abnormality indexes in the IDH prediction, and could be integrated with other imaging biomarkers together to construct an interpretable model for IDH prediction.
Several molecular biomarkers (e.g. TERT, 1p/19q co-deletion, MGMT) are also found to be related with the treatment selection and prognosis prediction. Lower grade glioma with TERT mutation is reported to have better prognosis, and TERT mutation is also a promising indicator for the treatment response of radiotherapy and temozolomide in primary glioblastoma multiforme (GBM) [21][22][23][24][25]. 1p/19q co-deletion is a strong predictor of better prognosis for patients with oligodendroglioma after radiotherapy or alkylating chemotherapy [26,27]. GBM patients with methylated MGMT are more sensitive to temozolomide and radiotherapy, resulting in better prognosis [28][29][30]. For predictions of these molecular markers, Radiomics models are usual ways with acceptable results, however, Radiomics features are also not intuitional and dependent on many factors (e.g. scanning image parameter, scanning machine type). In our study, several abnormal indexes were found with direct correlational relationship with these molecular markers, and combination of all indexes could display high correlations with them (r=0.619,0.615 and 0.480 respectively). The results indicate that the intrinsic molecular characteristics of tumor could be re ected by the structural abnormal pattern in glioma patients.
Finally, one limitation should not be ignored: the study is with limited data from single-center. Future studies should include more patients from multicenter to verify the usefulness of the novel structural indexes.

Conclusion
We propose a series of structural indexes for glioma patients to quantify the individual structural abnormal characteristics, which are found to correlate with tumor-speci c features such as WHO grade and molecular status as well as neurocognitive performance. The proposed indexes enhance the diagnostic information from conventional structural MRI, perhaps allowing for a more holistic assessment of disease severity for individual glioma patients and helping neurosurgeons to determine the personalized treatment strategies. Due to technical limitations, Table 2 is only available as a download in the supplementary les section. Table 3 The correlation value between individual structural abnormality index and molecular indicators.  Figure 2 Several examples of personalized GM/WM abnormality maps.