Evaluating angioarchitectural characteristics of glial and metastatic brain tumors with conventional magnetic resonance imaging

Abstract

Primary and metastatic brain tumors can overlap in traditional imaging features detected on preoperative conventional magnetic resonance imaging (MRI). The research objective was to determine whether morphological vascular characteristics present in routine preoperative imaging using traditional MRI sequences are predictive of primary versus metastatic brain tumors; secondarily to determine association of conventional and vascular-related imaging parameters with intraoperative blood loss, pathological invasion, and World Health Organization (WHO) tumor grade. A retrospective review analyzed 100 consecutive intracranial tumor surgeries, 50 WHO grade II-IV gliomas and 50 intracranial metastases. Two blinded expert readers independently evaluated preoperative MRIs, obtained via standard morphological imaging sequences, for adjacent or intra-tumoral arterial aneurysm, peritumoral venous ectasia, prominence, or engorgement (“aberrant peritumoral vessels”), and prominent intra-tumoral flow voids. Multivariate analysis was performed to develop models predictive of glioma and glioblastoma (GBM). Aberrant peritumoral vessels and prominent intra-tumoral flow voids were statistically significant predictors of glioma in univariate analyses (p = 0.048, p = 0.001, respectively) and when combined in multivariate analysis (OR = 5.23, p = 0.001), particularly for GBM (OR = 9.08, p < 0.001). Multivariate modeling identified prominent intra-tumoral flow voids and FLAIR invasion as the strongest combined predictors of gliomas and GBM. Aberrant peritumoral vessels and larger tumor volume predicted higher intraoperative blood loss in all analyses. No vascular-related parameters predicted pathological invasion on multivariate analysis. Aberrant peritumoral vessels and prominent intra-tumoral flow voids were predictive of gliomas, specifically GBM. These vascular characteristics, evaluated on routine clinical preoperative MRI imaging, may aid in distinguishing glioma from brain metastases and may predict intraoperative blood loss.

1. Introduction

Glioblastoma (GBM) is a World Health Organization (WHO) grade IV tumor, and represents the most common malignant primary brain tumor in adults [1]. Distinction of GBM from WHO grade II and III gliomas and metastatic brain lesions on preoperative magnetic resonance imaging (MRI) can help predict diagnosis, prognosis, and treatment planning prior to initial surgery. Distinguishing between primary, high-grade gliomas and aggressive metastatic lesions is obfuscated by overlap in traditionally evaluated MRI characteristics [2,3], particularly in cases with an unknown primary cancer site. Recently, some imaging features have been validated as predictors of poorer outcome in GBM, including periventricular location, involvement of the corpus callosum, eloquent location, large diameter, and associated edema [4]. While angioarchitectural anatomy has emerged as a possible predictive tool in classifying pathology and predicting tumor behavior, most studies have relied on extensive data analysis [5], specific quantitative non-routine sequences [6,7], or machine learning techniques [8,9] in order to identify associations between anatomy and clinical outcome.

This research investigated whether 3 vascular-related parameters present in routine clinical preoperative imaging using traditional MRI sequences are predictive of primary and metastatic brain tumors, specifically abnormal or prominent peritumoral venous drainage, associated arterial aneurysm, and T2 intratumoral vessels. These parameters were chosen based on the observation that GBM growth is associated with aberrant neovascularization [10], coupled with previously reported [11,12] and internal institutional observations that GBMs form delicate venous networks in the periventricular white matter (i.e. “medullary veins”). We hypothesized that using these vascular-related parameters in combination with conventional parameters would enhance the ability of routine MRI to distinguish between primary and metastatic brain tumors, thus increasing the utility of routine MRI in the clinical setting. Secondary outcomes included determining associations of conventional and vascular-related imaging parameters with clinical outcomes, including intraoperative blood loss, pathological invasion, and WHO tumor grade.

2. Materials and methods

2.1. Patient population

Our institutional review board waived consent as only retrospective review of medical records was performed. Following institutional review board approval, medical records at our institution were reviewed to identify 100 consecutive intracranial tumor surgeries performed by 2 surgeons for patients > 18 years from October 2013 to October 2015, including 50 consecutive intracranial metastases and 50 consecutive WHO grade II-IV glioma. Images were not reviewed at this point. Patients were excluded if they had a prior craniotomy at an outside hospital, had undergone a primary craniotomy at our institution before 2005, and when final pathology was nonspecific, unknown, or WHO grade I (for gliomas). If patients had multiple lesions (i.e. multifocal glioma or metastatic disease) or subsequent surgeries for residual or new disease, the first operative lesion was used for analysis.

2.2. Definition of variables

The electronic medical record (EMR) provided clinical information regarding presentation, surgical course, treatment, and clinical follow-up. Surgical resection was defined as subtotal resection (STR, <90%), near total resection (NTR, 90 to 95%), or gross total resection (GTR, 95% or greater). Pathology reports were reviewed for final diagnosis and presence of hemorrhage, invasion, or necrosis. Operative reports were reviewed for estimated blood loss (EBL). For lesions with multiple surgical resections, clinical and radiographic data were recorded from initial presentation (i.e. prior to first craniotomy). Patients with secondary GBMs were included, however data were included strictly from the first craniotomy and pathological diagnosis. Biopsied lesions were included only if the final neuropathological diagnosis was confirmed at our institution. Imaging features evaluated were identified on T2 Fluid attenuated inversion recovery (FLAIR), T2-weighted (T2), and T1-weighted (T1) images. All evaluated T1 sequences were obtained following the intravenous injection of a gadolinium-based contrast agent (GBCA).

2.3. Evaluation of preoperative MRI

Two blinded, expert readers, 1 neuroradiologist and 1 neurosurgeon/neuro-interventionalist, independently evaluated MRI scans. Both readers first analyzed a test set of 50 tumors (25 glioma, 25 metastatic), independent of the tumors included in the final analysis. Differences in the test analysis were reconciled via in-person discussion, and utilized to establish definitions of parameters.

Each reader was instructed to assess 3 conventional imaging features and 3 vascular-related imaging features for each patient. Conventional imaging parameters were based on previously validated measures and included (1) periventricular location, defined as enhancing tumor margin within 5 mm of the ventricle [4], (2) “invasion”, defined as T2 or FLAIR hyperintensity with indistinct tumor margins relative to brain and peritumoral edema, with asymmetric expansion of cortical and/or subcortical structures [13], and (3) significant associated but circumscribed FLAIR edema (>5 mm), suggesting mass effect rather than invasion [14]. Vascular-related imaging features assessed for each patient included (1) presence of a suspected intra-tumoral or associated arterial aneurysm, defined as a saccular appearing flow void on T2 images, measuring at least twice the caliber of the adjacent parent vessel; (2) peritumoral venous ectasia, prominence, or engorgement (collectively referred to as “aberrant peritumoral vessels”); and (3) prominent intra-tumoral flow voids, greater than approximately 30% of the parent vessel diameter. Reliability between readers was assessed by calculating the interclass correlation coefficient (ICC) for each imaging feature evaluated. ICC values < 0.4 were considered to indicate “poor” agreement between expert readers, values 0.4–0.6 were defined as “fair”, values 0.61–0.75 were defined as “good”, and values 0.76–1.0 were defined as “excellent”.

2.4. Statistical analysis

Primary outcomes included determining the association of conventional and vascular-related imaging parameters with pathological diagnoses of any glioma, GBM specifically, or metastases. Secondary outcomes included association of imaging findings with the presence of hemorrhage, invasion, and/or necrosis on pathological diagnosis, advanced WHO grade (glioma group), intra-tumoral hemorrhage on MRI at presentation, and surgical EBL. Statistical comparison of the pathology cohorts was performed using Fisher’s exact and Mann-Whitney tests as appropriate. Primary outcomes were analyzed using both univariate and multivariate (best forward selection) models using mixed-effects logistic regression, with the expert readers modeled as random effects and MRI characteristics as fixed effects. For secondary outcome analysis, WHO grade was similarly analyzed using mixed-effects ordinal regression. EBL was analyzed using fixed-effects linear regression (due to convergence issues) with MRI effects characterized as the average diagnosis of the two readers. Tumor volume and pathology were included in the multivariate analysis for EBL (in addition to all 6 radiographic parameters) to account for potential confounding bias. The remaining secondary outcomes were analyzed using mixed-effects logistic regression to create a model predictive of glioma. Significance was assessed at p < 0.05, with p-values adjusted for multiple comparison as necessary via the Benjamini-Hochberg procedure [15].

3. Results

3.1. Demographics and clinical characteristics of study groups

Table 1 shows the overall demographics of the 2 study groups. Average age was 53 years in the glioma group and 57 years in the metastasis group. Thirty-one patients (62%) of the glioma group had GBM, 12 (24%) WHO grade III, and 7 (14%) WHO grade II. Sixty percent of gliomas and 50% of metastatic lesions were left sided. There was a higher representation of temporal lobe location for gliomas (48% versus 14%, p < 0.001). The most common primary malignancy was lung adenocarcinoma, followed by melanoma. Metastatic lesions were smaller compared with gliomas (mean diameter 3.7 versus 4.4 cm, respectively, p < 0.001).

Table 1.

Demographics and clinical characteristics of glioma and metastatic patient cohorts.

		Glioma n = 50 (%)	Metastatic Lesion n = 50 (%)	p-value
Age (years)		53 ± 15	57 ± 14	0.215
Gender	Female	15 (30%)	15 (30%)	1
	Male	35 (70%)	35 (70%)
Side of Operative lesion	Left	30	25	0.422
Location	Frontal	17 (34%)	9 (18%)	<0.001
	Parietal	8 (16%)	16 (32%)
	Temporal	24 (48%	7 (14%)
	Occipital	0 (0%)	12 (24%)
	Brainstem	1 (2%)	0 (0%)
	Cerebellum	0 (0%)	6 (12%)
Mean Tumor Diameter (cm)		4.6 ± 1.6	3.7 ± 1.4	<0.001
Tumor Volume (mL)		37 ± 41	22 ± 18	0.044
Estimated Blood Loss (EBL) (mL) *		239 ± 114	177 ± 106	<0.001
EBL/Tumor Volume *		17.1 ± 30.6	20.3 ± 32.1	0.410
WHO Grade/Primary site	Grade II	7
	Grade III	12	NA
	Grade IV	31
	Lung adenocarcinoma		14
	Melanoma		13
	Breast		4
	Germ cell		3
	Lung squamous cell		3
	Uterine		2
	Renal clear cell		2
	Chordoma	NA	1
	Prostate adenocarcinoma		1
	Synovial carcinoma		1
	Colon adenocarcinoma		1
	Mesothelioma		1
	Thyroid adenocarcinoma		1
	Esophageal adenocarcinoma		1
	Lung large cell neuro-endocrine		1
	Lung small cell		1

Bold type indicates statistically significant (p < 0.05), WHO = World Health Organization.

^*= excludes biopsy (1) or cases with multiple lesions resected (6).

The majority of patients (63/100) underwent MRI examination the morning of surgery for surgical planning. MRI always included T1- and T2-weighted images, both performed following intravenous administration of a GBCA (95 patients). Five patients did not receive contrast due to medical allergy. Reported administration of GBCA ranged from 5 to 30 mL. On occasion, a separate T2-FLAIR sequence was obtained during the same imaging session, post administration of GBCA. T1-weighted images comprised both traditional 2-dimensional (turbo spin echo vs. fast spin echo, 90/100 patients) and 3-dimensional (spoiled gradient echo, 10/100 patients) techniques, resulting in greater parameter heterogeneity. Imaging parameters for T1 + GBCA (63 patients) included: TR = 23, TE = 2.302, matrix size = 252 × 252, pixel size = 0.71 mm × 0.71 mm, and slice thickness = 2.6 mm (SPGR sequence, 3D). The remaining patients (n = 37) had the following ranges for T1 + contrast parameters: TR = 7.62 – 1759, TE = 1.808 – 16.192, Matrix size = (224 – 384) × (173 – 302), pixel size = 0.43 × 0.43 – 1.094 × 1.094 mm, slice thickness = 1–5.5 mm. Imaging parameters for T2 sequences (58 patients) comprised the following parameters: TR = 3000, TE = 90, matrix size = 456 × 307, pixel size = 0.406 mm × 0.406 mm, and slice thickness = 3 mm. The remaining patients (n = 42) had the following ranges for T2 parameters: TR = 1400 – 8002, TE = 80 – 322, Matrix size = (224 – 512) × (192 – 454), pixel size = 0.375 × 0.375 – 1.016 × 1.016 mm, slice thickness = 1–5.5 mm. A total of 94 patients had preoperative FLAIR sequences, 8 of which were obtained on the pre-surgical planning study, with the following parameter ranges: TR = 4800 – 11,000, TE = 74 – 358.6, matrix size = (208–384) × (49–320), pixel size = 0.429 × 0.429 mm – 0.977 mm × 0.977 mm, slice thickness = 0.9 – 5.5 mm.

3.2. Vascular-Related and conventional MRI parameters

Table 2 shows results from review of imaging findings by the 2 independent readers. Two of the 3 conventional imaging features showed a higher representation in the glioma and GBM cohorts, including periventricular location (p = 0.004 and p = 0.005, respectively) and perilesional FLAIR hyperintensity suggestive of local invasion and/or infiltrative edema (p < 0.001 for both). For vascular-related parameters, both readers identified a statistically significant, higher rate of aberrant peritumoral vessels in the glioma group compared with the metastatic group in univariate analysis (p = 0.048), more pronounced when comparing GBM to metastasis (p = 0.027). Fig. 1 shows an example of an aberrant peritumoral vessel. Similarly, prominent intra-tumoral flow voids were significantly more common in the glioma group (p = 0.001) and the GBM subgroup (p < 0.001). An example of prominent intra-tumoral flow voids is shown in Fig. 2. There were no differences in the rates of associated arterial aneurysms. ICC values for all imaging features are shown in Table 2. Overall, the 2 readers showed fair agreement on 2 conventional and 2 vascular-related parameters, poor agreement on 1 vascular-related parameter (identification of aberrant peritumoral vessels) and good agreement on 1 conventional parameter (periventricular location).

Table 2.

Overview of vascular-related and conventional MRI parameters evaluated by each reader.

			Reader A		Reader B		p-value (Glioma vs. Met)	p-value (GBM vs. Met)	Interclass Correlation Coefficient
			Glioma n (%)	Met n (%)	Glioma n (%)	Met n (%)
Vascular-Related Parameters	Aberrant peritumoral vessels	Yes	32 (64%)	19 (38%)	15 (30%)	6 (12%)	0.048	0.027	0.35
		No	18 (36%)	31 (62%)	35 (70%)	44 (88%)
	Associated arterial aneurysm	Yes	5 (10%)	1 (2%)	1 (2%)	0 (0%)	0.275	0.209	0.43
		No	45 (90%)	49 (98%)	49 (98%)	50 (100%)
	T2 intra-tumoral vessels	Yes	41 (82%)	19 (38%)	38 (76%)	21 (42%)	0.001	<0.001	0.48
		No	9 (18%)	31 (62%)	12 (24%)	29 (58%)
Conventional Parameters	Periventricular location	Yes	35 (70%)	18 (36%)	35 (70%)	19 (38%)	0.004	0.005	0.67
		No	15 (30%)	32 (64%)	15 (30%)	31 (62%)
	Invasion	Yes	46 (92%)	1 (2%)	24 (48%)	4 (8%)	<0.001	<0.001	0.58
		No	4 (8%)	49 (98%)	26 (52%)	46 (92%)
	Associated edema	None	1 (2%)	0 (0%)	5 (10%)	1 (2%)	0.171	0.936	0.51
		<5 mm	16 (32%)	2 (4%)	1 (2%)	4 (8%)
		>5 mm	33 (66%)	48 (96%)	44 (88%)	44 (88%)

Bold type indicates statistically significant (p < 0.05), GBM = Glioblastoma multiforme, Met = metastatic lesion.

Fig. 1.

Aberrant peritumoral, periventricular vessels on a T1-post contrast axial image in a glioblastoma. Note the prominent deep medullary vein (white arrow) that is asymmetrically pronounced on the side of the tumor.

Fig. 2.

Prominent intra-tumoral flow voids on a T2 axial slice in a glioblastoma.

4. Multivariate modeling

Multivariate analysis was performed to create models predictive of glioma and GBM (Table 3). For glioma, the model was strongest with input from prominent intra-tumoral vessels (OR = 4.37) and perilesional FLAIR hyperintensity suggesting invasion (OR = 37.59). Similar results were obtained after creating a model to predict GBM.

Table 3.

Multivariate modeling for predicting gliomas and Glioblastoma multiforme.

			Predicting Glioma				Predicting GBM
	Variables	Significant Variable(s)	OR	CI (95%)	p-value	PP of glioma	OR	CI (95%)	p-value	PP of GBM
Vascular-Related Parameters	Aberrant peritumoral vessels; arterial aneurysms; T2 intratumoral vessels	T2 intratumoral vessels	5.23	1.99–13.69	0.001	65%	9.08	2.72–30.09	<0.001	53%
Conventional Parameters	Periventricular; invasion; edema	Invasion	44.24	12.99–150.73	<0.001	93%	31.93	8.25–123.62	<0.001	87%
Combination	Aberrant peritumoral vessels; arterial aneurysms; T2 intratumoral vessels; periventricular; invasion; edema	T2 intratumoral vessels	4.37	1.54–12.35	0.006	95%	5.41	1.56–18.72	0.009	89%
		Invasion	37.59	10.92–129.43	<0.001		21.24	5.32–84.8	<0.001

Bold type indicates statistically significant (p < 0.05), OR = Odds ratio, CI = confidence interval, GBM = Glioblastoma multiforme, PP = predicted probability.

4.1. Estimated blood loss, histopathological invasion

Of the secondary outcomes investigated, aberrant peritumoral vessels and associated arterial aneurysm were predictive of a higher EBL (138 mL, p < 0.001; 221 mL, p = 0.007, respectively); intra-tumoral flow voids predicted both histopathological invasion (OR = 3.16, p = 0.017) and a higher absolute EBL (111 mL, p < 0.001) (Table 4). Multivariate analysis for pathological invasion and EBL showed that radiographic suggestion of invasion from FLAIR sequences best predicted pathological invasion, while aberrant peritumoral vessels and higher tumor volume best predicted higher EBL (Table 5). Pathology (metastasis versus glioma) was included in the multivariate model for EBL but was not significant. Predicted probability values are listed for both multivariate models. Thus, the presence of radiographic invasion yielded a 75% prediction of pathologic invasion, compared to 39% without any. Similarly, using the average tumor volume across the whole cohort (29 mL), the presence of aberrant peritumoral vessels predicted an EBL of 262 mL, compared to 178 mL without this parameter.

Table 4.

Univariate analysis for secondary outcomes.

	Pathology Report				Hemorrhage on MRI	EBL from surgery*
	Higher WHO Grade	Hemorrhage	Necrosis	Invasion
Aberrant peritumoral vessels	N/A	N/A	N/A	N/A	N/A	138 mL p < 0.001
Associated arterial aneurysm	N/A	N/A	N/A	N/A	N/A	221 mL p = 0.007
T2 intra-tumoral vessels	N/A	N/A	N/A	p = 0.017 (OR = 3.16)	N/A	111 mL p < 0.001
Periventricular	N/A	N/A	N/A	p = 0.033 (OR = 2.87)	N/A	105 mL p < 0.001
Invasion	p = 0.001 (OR = 9.74)	p = 0.049 (OR = 0.36)	N/A	p = 0.004 (OR = 4.63)	0.096 (OR = 0.37)	59 mL p = 0.045
Edema > 5 mm	p = 0.034 (OR = 0.14)	N/A	p = 0.049 (OR = 8.94)	N/A	N/A	N/A

Bold type indicates statistically significant (p < 0.05), OR = Odds ratio, WHO = World Health Organization, MRI = magnetic resonance imaging, EBL = estimated blood loss, N/A = not associated. * = excludes biopsy (1) or cases with multiple lesions resected (6).

Table 5.

Multivariate modeling for secondary outcomes of estimated blood loss and pathological invasion.

Outcome	Significant radiographic variables	OR	CI (95%)	p-value	Predicted probability values
Pathological invasion	Invasion	4.63	1.65–12.98	0.004	75% glioma if “+”	39% glioma if “−”
Estimated blood loss*	Aberrant peritumoral vessels	–	13.3–138	0.018	262 mL if “+”, average tumor volume1	178 mL if “−”, average tumor volume1
	Tumor Volume	–	0.6–1.94	<0.001

OR = odds ratio; CI = confidence interval. * = excludes biopsy (1) or cases with multiple lesions resected (6). ¹Using an average tumor volume of 29 mL.

5. Discussion

5.1. Vascular-related features predict gliomas, especially GBM

In this research, we present a unique analysis of WHO grade II, III, and IV gliomas and metastatic lesions using routine MRI sequences, with the goal of distinguishing between primary and secondary brain tumors using both conventional and vascular-related imaging parameters. We identified two vascular-related features predictive of gliomas, and specifically of GBM: aberrant peritumoral vessels (venous ectasia, prominence, and engorgement) and prominent intra-tumoral flow voids. Angiographically-proven associated arterial or intra-tumoral aneurysms, likely related to aberrant arteriovenous shunting, were rare in our series (1 total detected from 7 angiograms). This is consistent with the approximately only 20 cases reported in association with GBM [16], precluding meaningful statistical analysis in our series and overall.

There has long been interest in identifying subtle abnormalities in cerebral vasculature related to tumor pathology. In 1974, Hooshmand et al. highlighted “deep medullary veins”: long, fine vessels draining the deep cerebral white matter, oriented perpendicular to the long axis of the lateral ventricle and atrium, with convergence at the superolateral angle of the ventricle and drainage into subependymal veins [11,12]. They noted that in neoplastic, inflammatory, or vascular pathology affecting the deep white matter, deep medullary veins appeared earlier in angiography, with larger diameters (>200 μm), longer lengths, increased tortuosity, and multiplicity or clustering [11]. Another early study of both angiography and CT attempted to differentiate high grade gliomas from metastasis, with early venous drainage, angiographic “stain”, and abnormal vessels as angiographic markers for GBM [17]. More recent analyses have focused on differentiating metastatic lesions from GBM and other primary brain tumors using routine MRI sequences. Baris et al. found that metastatic tumors had less midline shift, lower tumor volume, and more mass edema than primary brain tumors (including GBM) [2], similar to our results showing that metastatic lesions were significantly smaller than gliomas. Based on our results, both aberrant peritumoral vessels and T2 intra-tumoral vessels are more likely to be present in gliomas and GBM than in metastatic disease. Via multivariate modeling, the latter, combined with FLAIR suggestions of invasion, produce the strongest predictors of both gliomas and GBM. These vascular features have clinical significance, as they are associated with higher EBL during surgery. In future, this information could be used to develop operative expectations for various neoplasms.

5.2. Measuring angiogenesis on imaging

Angiogenesis plays an important role in neovascularization and maturation of gliomas, GBM, and metastatic lesions [18]. Via upregulation of growth factors, such as vascular endothelial growth factor, angiogenesis enables supply of nutrients and oxygen and is essential to tumor growth, particularly for GBM. Advanced imaging techniques have been applied in an effort to measure tumor angiogenesis. For example, perfusion weighted imaging (PWI) sequences allow for voxel-based measurements across enhancing tumor tissue and quantification of relative cerebral blood volume [19]. Stratification of patients into different subgroups based on PWI features has been demonstrated to accurately predict response to anti-angiogenic therapies [20]. Even more sophisticated sequences can estimate “microvascular vessel caliber” via differences between sensitivity to gradient echo (GRE) and spin echo (SE) MRI compared to magnetic susceptibility, with GRE signals present in both macro- and microscopic vessels, and SE signals specific to microscopic vessels [5,21,22]. However, these sequences are not performed in most routine MRI studies, especially in the community setting. Recently, Puig et al. analyzed “macrovascular networks” in GBM using T1-contrasted MRI and found that a more robust network correlated with poor prognosis [23]. Although precise computation of the network requires perfusion imaging and analysis, the underlying phenomenon is similar to our findings that significant intra/peritumoral vessels are predictive of glioma over metastasis.

5.3. Independent evaluation of preoperative MRIs

While achieving statistical significance in multivariate modeling, some of the results produced broad confidence intervals (see invasion, Table 3). This was due to a difference in sensitivity and specificity between the 2 readers. For example, in the associated arterial aneurysm category, Reader A detected 6 suspected arterial aneurysms, compared to just 1 from Reader B, producing a fair ICC. A second review of these cases, including the Readers, a third party adjudicator and a reiteration of definitions resulted in both Readers selecting the same predictions for all 6 patients. For cases that Reader A had concern for associated arterial aneurysm, Reader B interpreted these to be either local venous ectasia or prominent intra-tumoral flow voids. Thus, there was disparity between the 2 readers attributable to differences in training and natural variability in interpreting imaging data. One reader was a dual trained cerebrovascular neurosurgeon with an appointment in the radiology department, and the other was a board certified neuroradiologist; both were located at an academic setting, and recapitulate the interdisciplinary process employed at our institution for the care of patients with brain tumors. In the absence of vascular imaging (i.e. formal cerebral angiogram, MR or computed tomography angiography), there is no definitive answer as to whether or not a “vascular” parameter is present. Serial training sets with angiographic data to support predictions and a prospective validation of these features in a series of ambiguous tumors prior to tissue diagnosis would increase confidence in their predictive value, narrow the confidence intervals, and improve the ICCs.

5.4. Limitations

Due to its retrospective design, this study has limitations, including possible selection bias. In response, all consecutive patients meeting inclusion criteria were included, without reviewing imaging at the time of inclusion, and expert readers were blinded to all pathological and radiological reports. Patients with prior surgical debulking or surgical resection were excluded, in an effort to isolate the initial presentation of a tumor and its vascularity. The majority of ICC values fell into the “fair” category, suggesting that additional expert reader input or more strict imaging criteria might be useful in enhancing reliability. Challenges in obtaining high ICC have been previously identified in neuroradiological studies of “routine” diagnoses [20]. Despite this limitation, sample size was sufficient to demonstrate statistically significant differences between groups. It is also possible that confounding bias played a role, particularly in the secondary outcome analysis. This was adjusted for by including variables such as tumor volume and pathology that seemed to have separate statistical relationships to the secondary outcome of interest. Another potential confounder is the variability of the acquired imaging sequences utilized to obtain the evaluated parameters. However, this variability is representative of routine clinical imaging for the patient populations regularly treated at our institution. As such, this variability demonstrates the durability of the reported imaging findings across multiple imaging parameters within a real-world clinical setting.

6. Conclusions

Two vascular features (aberrant peritumoral vessels and prominent intra-tumoral flow voids), seen on non-vascular imaging, are associated with glioma and GBM rather than brain metastases. These results suggest that, in addition to traditional MRI findings, these vascular characteristics can aid in differentiating glioma - from brain metastasis, and may also be predictive of intraoperative blood loss.