Diffusion Histology Imaging Combining Diffusion Basis Spectrum Imaging (DBSI) and Machine Learning Improves Detection and Classification of Glioblastoma Pathology

This article is featured in Highlights of This Issue, p. 5269

Glioblastoma (GBM) is the most common primary malignant brain tumor in adults (1). It is estimated that 13,140 new GBM cases will be diagnosed during 2020 in the United States (1). Despite extensive multimodality treatment, which includes surgical resection, chemotherapy, and radiation, patients with GBM exhibit a dismal 5-year survival rate of 6.8% (1). Histologically, GBMs are characterized by increased cellularity, vascular proliferation, necrosis, and infiltration into normal brain parenchyma (2). Currently, the histopathologic complexity of GBM cannot be fully appreciated without microscopic examination of tumor specimens.

Gadolinium (Gd)-enhanced T1-weighted (T1W) MRI is the standard clinical imaging modality for detection, surgical planning, and evaluation of GBM treatment response (3–6). Contrast enhancement in T1W images (T1WI) is clinically interpreted as a measure of tumor burden and is widely used as the target for surgical resection (6). However, due to the infiltrative nature of GBM, tumor cells are known to extend well beyond the area of contrast enhancement (3). After treatment, contrast enhancement is not diagnostically specific for GBM, because it reflects not only increased Gd leakage because of angiogenesis induced by malignant tumors, but also the blood–brain barrier disruption triggered by other factors, including radiation effects and ischemia (7–9). Therefore, Gd enhancement neither accurately measures tumor burden nor specifically reflects various pathologic changes.

Conventional MR sequences such as T2-weighted imaging (T2W) and fluid-attenuated inversion recovery (FLAIR) imaging have also been employed to localize nonenhancing tumor to complement Gd-enhanced T1WIs. The combination of these imaging sequences was adopted into the Response Assessment in Neuro-Oncology (RANO) (3). However, precise quantification of increase in T2W/FLAIR image signal intensity remains difficult. Differentiating nonenhancing tumor from other causes of increased T2W or FLAIR image signal intensity, such as edema, radiation effect, ischemic injury, postoperative changes, or other treatment effects, continues to challenge clinicians.

In addition to conventional T1W and T2W imaging, diffusion-weighted imaging and the derived apparent diffusion coefficient (ADC) have been employed to detect and assess tumor cellularity in many cancers based on the hypothesis that increased tumor cellularity restricts diffusion, which decreases ADC values (10). ADC has been shown to decrease with increasing glioma grade (11) and applied to characterize the infiltrative pattern of recurrent tumor after treatment (12). However, ADC loses specificity and sensitivity in the presence of coexisting necrosis (increased ADC), tumor infiltration (decreased ADC), and/or vasogenic edema (increased ADC) that complicate local brain diffusion characteristics. The combination of multiple MR sequences falls short in predicting the complex and heterogeneous GBM tumor microenvironment. In addition, the gold standard of surgical biopsy carries risk. Thus, the development of noninvasive alternatives to decipher the complex GBM tumor histopathology remains an urgent need so that clinicians can make rational decisions about continuing, stopping, or changing treatments.

Diffusion basis spectrum imaging (DBSI) utilizes a data-driven multiple-tensor modeling approach to differentiate coexisting morphologic features resulting from tumor pathologies or other attributes within an image voxel. We have previously demonstrated that DBSI quantifies tissue injury in an array of central nervous system (CNS) disorders including multiple sclerosis (13–15), cervical spondylotic myelopathy (16), and epilepsy (17). In this study, we demonstrate both Gd-enhanced T1WIs and hyperintense FLAIR areas contain a spectrum of DBSI-derived morphologic signatures in GBM. Using a modified DBSI algorithm to separate inflammation from tumor cellularity, we show DBSI-derived restricted isotropic diffusion fraction positively correlated with tumor cellularity in GBM specimens. Finally, to improve the performance of a DBSI-based cancer detection, we developed a robust diffusion histology imaging (DHI) approach by combining a machine learning algorithm with DBSI metrics to accurately identify and classify various histopathologic components of GBM.

This study was approved by the Institutional Review Board of the Washington University School of Medicine (St. Louis, MO) and conducted in accordance with the Declaration of Helsinki. Written informed consents were obtained from all participants. The goal was to develop a reliable and consistent neuroimaging outcome measure to accurately classify high cellularity tumor, tumor necrosis, and tumor infiltration in high-grade glioma. The inclusion criteria were: (i) adult subjects scheduled for brain tumor resection at the Washington University School of Medicine (St. Louis, MO), (ii) subjects who had not received radiotherapy or chemo therapy, and (iii) subjects whose resected tumor specimen were sufficiently large for ex vivo MRI in addition to that required for clinical diagnosis. Sixteen newly diagnosed adult patients with GBM (Supplementary Table S1) were recruited for in vivo (n = 3) and ex vivo (n = 13) MRI studies from June 2015 to January 2017. Eighteen patients with GBM suspicion were recruited for eligibility assessment from June 2015 to January 2017 (Supplementary Fig. S1). Two patients were excluded because of data problem and pathologic assessment of non-GBM (anaplastic ependymoma), respectively. After exclusion, 16 newly diagnosed adult patients with GBM without any previous treatments were included for in vivo (n = 3) and ex vivo (n = 13) MRI studies. The patient characteristics are summarized in Supplementary Table S1.

Nineteen resected specimens from 13 patients with GBM underwent multi-slice/section ex vivo MRI and histologic examinations (Supplementary Fig. S1). The average size of the specimens was 8 ± 4 mm. Among the 19 specimens, at least one specimen was obtained from each individual; for three individuals, specimens were taken from two sectors in the tumor; and from one individual, specimens were taken from four tumor sectors. Each tissue specimen contains multiple image slices and histology sections for MRI-histology coregistration and quantification. Multi-slice/section DBSI and hematoxylin and eosin (H&E) revealed that patterns of GBM pathologies were similar throughout the thickness of all specimens with the exception of two specimens from one individual, in which H&E patterns were distinct in two sections. Thus, we analyzed a total of 21 DBSI-H&E–matched sections from 19 specimens.

After resection, specimens were immediately fixed in 10% formalin in PBS (pH = 7.4) at room temperature for at least 48 hours (Supplementary Fig. S2A) and then transferred to PBS. PBS was changed every 2 days for a total of 2 weeks before the experiment. The specimens were embedded in agar gel for MRI and then analyzed using DBSI and diffusion tensor imaging (DTI).

Specimens were formalin-fixed at the time of collection and then agarose gel–embedded (Supplementary Fig. S2B) before being examined using a 4.7-T Agilent MR Scanner (Agilent Technologies) with a home-made circular surface coil (1.5-cm diameter; Supplementary Fig. S2C). A multi-echo spin-echo diffusion-weighting sequence with 99 diffusion-encoding directions (maximum b-value at 3,000 seconds/mm²) was employed to acquire diffusion-weighted images (DWIs) with a 0.25 × 0.25 mm² in-plane resolution and 0.5-mm thickness. The imaging parameters were as follows: repetition time (TR), 1,500 milliseconds; echo time (TE), 40 milliseconds; time between application of gradient pulse, 20 milliseconds; diffusion gradient on time, 8 milliseconds; slice thickness, 0.5 mm; field-of-view, 24 × 24 mm²; data matrix, 96 × 96; number of average, 1; and in-plane resolution, 0.25 × 0.25 mm². Total acquisition time was approximately 4 hours. MR images were zero-filled to 0.125 × 0.125 mm² in-plane resolution for DBSI and DTI analyses.

The specimens underwent sequential sectioning at 5-μm thickness. Sections were individually stained with H&E and glial fibrillary acidic protein (GFAP). Histology slides were digitized using NanoZoomer 2.0-HT System (Hamamatsu Photonics) with 20 × objectives for analyses. Each tissue specimen contains multiple image slices and histology sections for MRI-histology coregistration and quantification.

We developed a procedure involving down-sampling histologic images and coregistering MRI with histologic images to enable the voxel-wise correlation between tumor cellularity and MRI-derived surrogate marker of cellularity, for example, DTI-derived ADC and DBSI-derived restricted diffusion fraction. Detailed analysis was performed as described in Supplementary Materials and Methods. Briefly, specimens were sectioned and stained after ex vivo MRI to acquire H&E and GFAP images. High resolution histology images were down-sampled to match DBSI/DTI resolution (125 × 125 μm²) to enable direct comparison between DBSI/DTI and histologic images. Each down-sampled histology image voxel contained 272 × 272 original image pixels. A two-dimensional thin plate spline (TPS) coregistration method was adopted using 30 manually picked landmarks.

A 3-T Siemens TIM Trio (Erlanger) with a 32-channel head coil was used for all in vivo images. Axial DWIs covering the whole brain were acquired using a multi-b-value diffusion weighting scheme (99 directions; maximum b-value, 1,500 seconds/mm²) with the following parameters: TR, 10,000 milliseconds; TE, 120 milliseconds; field of view (FOV), 256 × 256 mm²; slice thickness, 2 mm; in-plane resolution, 2 × 2 mm²; and total acquisition time, 15 minutes. Eddy current and motion artifacts of DWI were corrected before susceptibility induced off-resonance field was estimated and corrected. Conventional MRI sequences including Gd-enhanced T1W image, anatomic 3D magnetization prepared rapid acquisition of gradient echo (MPRAGE) image, T2W image, and T2W FLAIR image were performed per standard clinical protocol.

Pathologic examination following stereotactic biopsy or surgical resection plays a vital role in current clinical decision-making for the management of patients with GBM, based on the neuropathologist's recognition of morphologic signatures reflecting tumor cells and changes in the microenvironment, including treatment effects, which are characteristics missed by current MRI biomarkers. To address this critically important unmet need, we developed DBSI, which utilizes a data-driven multiple-tensor modeling approach to disentangle pathology and structural profiles within an image voxel (13, 18–22). Although DBSI-derived structural metrics distinguish and quantify various tissue pathologies in an array of CNS disorders (13, 16, 19, 23, 24, 25), the ability of DBSI to detect tissue microstructure alone is insufficient to accurately identify the underlying GBM pathologies of high tumor cellularity, tumor necrosis, and tumor-infiltrated white matter. We, thus developed a novel DHI approach, which applies machine/deep learning algorithms (26, 27) using DBSI structural metrics as input classifiers to accurately model underlying GBM pathologies.

DBSI models brain tumor diffusion-weighted MRI signals as a linear combination of discrete multiple anisotropic diffusion tensors and a spectrum of isotropic diffusion tensors:

Here, b_k is the kth diffusion gradient. S_k/S₀ is the acquired diffusion-weighted signal at direction of b_k normalized to nondiffusion-weighted signal. N_Aniso is number of anisotropic tensors to be determined. φ_ik is the angle between diffusion gradient b_k and principal direction of the ith anisotropic tensor. |$| {\overrightarrow {{b_k}} } |$| is b-value of the kth diffusion gradient. λ_||i and λ_⊥i are axial and radial diffusivity of the ith anisotropic tensor under the assumption of cylindrical symmetry; f_i is signal intensity fraction of the ith anisotropic tensor. a and b are low and high diffusivity limits of isotropic diffusion spectrum. f(D) is signal intensity fraction at isotropic diffusivity D.

On the basis of our ex vivo MRI and histologic analyses of resected specimens, the following isotropic diffusion profiles have been established on the basis of diffusivity. We observed that highly restricted isotropic diffusion (0 ≤ D ≤ 0.2 μm²/milliseconds) was associated with lymphocytes, restricted isotropic diffusion (0.2 < D ≤ 0.8 μm²/milliseconds) was associated with high tumor cellularity in GBM, and hindered isotropic diffusion (0.8 < D ≤ 2 μm²/milliseconds) was associated with tumor necrosis. For in vivo human subjects, the in vivo diffusivity profile can be estimated by extrapolating ex vivo diffusivity based on the temperature difference: highly restricted diffusion fraction (0 ≤ D ≤ 0.2 μm²/milliseconds; not affected by temperature), restricted isotropic diffusion (0.2 < D ≤ 1.5 μm²/milliseconds), and hindered isotropic diffusion (1.5 < D ≤ 2.5 μm²/milliseconds). Further detailed information can be found in the Supplementary Materials and Methods.

DBSI provides a simple tensor expression to visualize morphologic features resulting from tumor formation and nontumor entities appearing indistinct to tumor by conventional MRI. For example, in an image voxel where normal white matter tracts and gray matter are coexisting with the presence of tumor cells. The tensor function of anisotropic and isotropic tensors will not change comparing with the normal tissues, with the exception of changes resulting from the presence of tumor cells. If necrosis is present in tumor containing image voxels, it would require multiple isotropic diffusion tensors, such as restricted (modeling tumor cells) and hindered (modeling necrosis) isotropic tensors, to completely model the diffusion-weighted signals. The different tensor expressions of individual image voxels thus bear morphologic signatures of underlying pathology. In the case where tumor cells happen to also damage white matter tracts resulting in, say, axonal injury and demyelination, the isotropic tensors within this image voxel will remain the same but now anisotropic diffusion tensor will exhibit decreased axial diffusivity and increased radial diffusivity. It is the sensitivity of diffusion-weighted MRI signal to the microstructural changes in the scale up to a 10-μm range (depending how one adjusts diffusion-weighting condition) that allows DBSI to more precisely reflect morphologic changes resulting from tumor presence or other pathologic conditions. By taking the advantage of this feature of DBSI as inputs of machine learning algorithms, we created DHI to recapitulate histopathologic analysis using MRI.

We used Spearman rank correlation to measure strengths of monotonically increasing or decreasing associations between histology and MRI cellularity measurements. Statistically significant results were determined at a predetermined alpha level of 0.05.

To construct a machine learning classifier for histopathologic prediction, support vector machine (SVM) with polynomial kernel algorithm was adopted using a package by Scikit-learn in Python (28). We performed a linear coregistration between MRI and corresponding neuropathologist-classified H&E images to label each image voxel with the gold standard of pathology. A total of 21 sections from 19 brain tumor specimens (6,605 image voxels) were analyzed to determine DBSI and DTI metric profiles of each image voxel. Image voxels from four randomly selected sections were used for testing and the voxels from remaining sections were used for training. For cross-validation, a total of 1,000 distinct training–test group pairings were run to prevent selection bias. In addition, to address potential internal correlations of voxels from same patients, we performed 500 random splits that assigned voxels from different patients into training and test datasets. Mean values and 95% confidence intervals were calculated.

Confusion matrices were calculated to illustrate the specific examples of tumor pathologies where predictions were discordant with pathologist-identified pathologies. We evaluated overall classification accuracy of testing voxels as well as true positive rate, true negative rate, and positive predictive value of the model prediction. ROC and precision-recall curves were calculated using a one-versus-rest strategy to assess model discrimination for each tumor pathology. AUCs and F₁ scores were calculated to compare the relative performance of DHI with pathologist-identified pathologies.

Among the 16 patients, 11 were male and five were female. The mean age at diagnosis was 61.1 ± 14.2 years. Pathologic analysis of tumor specimens confirmed isocitrate dehydrogenase wild-type GBM in all 16 patients (Supplementary Table S1).

We performed clinical MRI and DBSI on three patients with GBM. Representative Gd-enhanced T1W, FLAIR, T2W, DBSI, and ADC images were obtained from a 79-year-old male patient with a right temporal GBM (Fig. 1A). We outlined Gd-enhanced and non-enhanced T1W regions to compare the underlying DBSI metrics in these regions, overlaid on MPRAGE-T1WIs (Fig. 1A). On the basis of our previous DBSI applications, we predicted that DBSI metrics of restricted fraction, hindered fraction, and anisotropic fraction would be seen in regions/voxels containing high tumor cellularity, necrosis, and fiber-like structures (neuronal fibers or extracellular matrix fibers), respectively. Strikingly, DBSI metrics of restricted fraction (red), hindered fraction (blue), and anisotropic fraction (green) were entangled in both Gd-enhancing and nonenhancing regions (Fig. 1A; DBSI). Specifically, hyperintensity of restricted fraction was widespread in Gd nonenhancing region where histology is typically considered to be necrosis, indicating the potential high tumor cellularity in this region that could significantly challenge the current clinical standard.

To determine whether specific DBSI structural metrics are enriched in particular clinical MRI sequences, we generated histograms of DBSI metrics from Gd-enhancing, nonenhancing, and FLAIR hyperintense lesions from all three patients (Fig. 1B). The common feature among the three GBM cases was the consistent presence of the three DBSI metrics in all clinical MRI-defined lesions. Qualitatively, the pattern of DBSI metric distributions did not appear to be unique for Gd-enhancing, nonenhancing, or FLAIR hyperintense regions of GBM tumors, suggesting that these clinical MRI-defined regions harbor mixed pathologies.

As shown above (Fig. 1), in vivo DBSI restricted, hindered, and anisotropic fractions were highly overlapping in MR lesions of GBM. To definitively determine relationships between DBSI metrics and GBM pathologies, we examined ex vivo DBSI metrics in histologically identified regions of high tumor cellularity, tumor necrosis, and tumor infiltration in 19 surgically resected specimens. We performed a TPS coregistration on specimens correlating diffusion-weighted images with H&E and GFAP cellularity maps (Fig. 2A and B) to allow voxel-to-voxel correlation of histology (H&E- and GFAP-positive area ratio maps) with ADC and DBSI-restricted fraction maps (Fig. 2C). We randomly selected 50 voxels from down-sampled H&E images (Fig. 2A, red squares) and mapped them to the coregistered GFAP, MRI-metric maps for voxel-based correlation. Of 20 specimens, 15 underwent MRI-H&E and nine underwent MRI-GFAP correlation analyses. The rest were excluded because of unmatched sectioning planes.

Spearman rank correlation for selected voxels from all specimens (Fig. 2C) was used to assess the general performance of imaging biomarkers for cellularity in tumor samples. Restricted fraction correlated with H&E (r = 0.53; P < 0.0001) and GFAP–positive areas (r = 0.66; P < 0.0001). In contrast, ADC failed to correlate with H&E (r = −0.078; P = 0.04) or GFAP (r = −0.055; P = 0.28). In addition, DBSI isotropic ADC showed slightly negative correlation with H&E cellularity (r = −0.16; P < 0.0001) and no correlation with GFAP cellularity (r = 0.008; P = 0.88).

To definitively determine relationships between DBSI metrics and GBM pathologies, we examined ex vivo DBSI metrics in histologically identified regions of high tumor cellularity, tumor necrosis, and tumor infiltration in 19 surgically resected specimens. A representative tumor specimen (10.1 × 8.7 mm²) from a 77-year-old female patient demonstrates the relationship between DBSI metrics and tumor pathologies (Fig. 3). Hyperintense DWI (i.e., hypointense ADC) defined voxels did not correspond to H&E measures of cellularity (Figs. 2, 3A and B). By coregistering DWI with histologic images, representative voxels from regions of high tumor cellularity (Fig. 3B, H&E and GFAP; red square), tumor infiltration (Fig. 3B, H&E and GFAP; green square), and tumor necrosis (Fig. 3B, H&E and GFAP; blue square) were selected to perform isotropic diffusion spectrum analysis (Fig. 3C). These representative image voxels of H&E and GFAP were enlarged and displayed for validating the histologic findings (Fig. 3B). In general, the isotropic diffusion spectrum analysis indicated three distinct clusters of diffusion signatures. Specifically, high tumor cellularity areas exhibited peaks at highly restricted and restricted diffusion regions of DBSI isotropic diffusion spectrum, infiltrated white matter exhibited peaks at the same locations as high cellularity with varying intensities, and tumor necrosis exhibited highly restricted and hindered diffusion peaks (Fig. 3C). On the basis of the diffusion spectrum analysis from each image voxel, the diffusion tensor fraction maps could be derived. Although these DBSI metrics were overlapping in these tumor pathologies (Figs. 1A, B and 3C), DBSI restricted and hindered fraction maps qualitatively resembled areas of high tumor cellularity and tumor necrosis, respectively, as identified by a neuropathologist (Fig. 3D).

Through image coregistration, MRI voxels corresponding to pathologically verified areas of high tumor cellularity, tumor necrosis, and tumor infiltration were identified. Image voxel values of DTI (Fig. 4A) and DBSI (Fig. 4B) metrics are presented to demonstrate the distinctions and similarities among these identified tumor pathologies. A multi-parametric examination on the basis of restricted fraction, hindered fraction, and isotropic fraction separated the three pathologically distinct entities (Fig. 4C), suggesting analysis based on multiple DBSI metrics could potentially better distinguish among these pathologies rather than single DBSI metrics alone.

We, thus developed DHI by incorporating a supervised SVM algorithm with modified (incorporating a distinction between inflammation and tumors) DBSI-derived structural metrics as classifiers to construct predictive models to distinguish among different tumor histopathologies. We trained and validated the predictive model on image voxels from 17 of the 21 GBM specimen sections. The established model was applied to image voxels from four remaining GBM specimen sections to predict distributions of high tumor cellularity (Fig. 4D, red), tumor necrosis (Fig. 4D, blue), and tumor infiltration (Fig. 4D, green) with 96.2% overall accuracy (n = 1,963). DHI correctly predicted 97.2%, 96.6%, and 91.8% of the image voxels as high tumor cellularity, tumor necrosis, and tumor infiltration, respectively.

A comparison between DHI and DTI-SVM was performed using confusion matrices. The DHI (Fig. 4E) approach demonstrated better prediction accuracies for tumor pathologies when compared with DTI-SVM results (Fig. 4F). We also performed ROC and precision-recall curves analyses for each tumor pathologic feature (Fig. 4E and F). DHI indicated greater ROC and precision-recall AUC values for all the pathologic features than DTI-SVM did.

From four DHI test specimens, we randomly selected four DWI voxels from each specimen (Fig. 5A–D) and used corresponding histology as validation. This was achieved by tracking each voxel back to the coregistered down-sampled histologic images to compare DHI-predicted pathologies with gold standards. We observed high predictive performance of DHI on individual specimens.

DHI predicted 94.3% of high tumor cellularity areas (Fig. 5A, red), 97.3% of necrotic areas (Fig. 5A, blue), and 82.1% of tumor infiltration areas (Fig. 5A, green) in a 77-year-old female patient specimen (B122). Corresponding H&E image tiles (i.e., voxels of down-sampled histology images) verified the randomly selected DHI predictions of high tumor cellularity (Fig. 5A, a and b), tumor infiltration (Fig. 5A, c), and tumor necrosis (Fig. 5A, d). The second test specimen from a 54-year-old male patient (B95) exhibited true prediction rates of 98.0% and 93.3% for high tumor cellularity voxels (Fig. 5B, red) and tumor necrosis voxels (Fig. 5B, blue), respectively, as validated by corresponding H&E image tiles of high tumor cellularity (Fig. 5B, a and b) and tumor necrosis (Fig. 5B, c and d). The third specimen from a 47-year-old female patient (B128) was also assessed, demonstrating that DHI-predicted voxels of high tumor cellularity were 99.0% accurate (Fig. 5C, a–d). In the fourth test specimen, from a 57-year-old female patient (B94), DHI correctly predicted 100% of tumor infiltration voxels (Fig. 5D). Corresponding H&E image tiles all indicated tumor infiltration patterns.

We ran 1,000 random training/validation and test split pairings to address possible selection bias resulting from the use of specific test samples. The mean accuracy of DHI was 89.6% compared with 76.7% of DTI-SVM. Mean true prediction rates of DHI for high tumor cellularity, tumor necrosis, and tumor infiltration were 87.5%, 89.0%, and 93.4%, respectively (Table 1). In contrast, mean true prediction rates of DTI-SVM were 76.7%, 62.3%, and 97.9%, respectively (Supplementary Table S2). In addition, DHI showed much better overall precision-recall performances, with mean F₁ scores of DHI being 0.917, 0.823, and 0.876 for three tumor pathologies. We further performed ROC analyses using one-versus-rest strategy to test how well our classifiers distinguish one tumor pathology from others (e.g., infiltration vs. noninfiltration). The ROC analysis results revealed that DHI had great performance on distinguishing these three tumor pathologies, with mean AUC values of 0.975, 0.989, and 0.951 for high tumor cellularity, tumor necrosis, and tumor infiltration, respectively.

To address the potential internal correlations from voxels of the same patients in training/validation and test datasets, we performed 500 random training/validation and test splits that assigned voxels from different patients into training/validation and test datasets. As expected, the accuracy was slightly lower, with 87.1% compared with 89.6% from the sample-wise split method. Similarly, the AUCs (high tumor cellularity, 0.960; tumor necrosis, 0.938; and tumor infiltration, 0.989) and F₁ scores (high tumor cellularity, 0.893; tumor necrosis, 0.816; and tumor infiltration, 0.816) were also slightly lower but comparable with results from sample-wise split method (Table 1), indicating the consistency and generalization of DHI.

The standard of care for GBM involves surgical resection, followed by radiotherapy with concurrent and adjuvant chemotherapy. Histologic assessment of tumor cellularity, necrosis, and infiltration plays a vital role in the clinical decision-making for the management of patients with GBM. The current gold standard of pathologic examination following stereotactic biopsy or surgical resection (29) carries potential risks (30). In some occasions, inconclusive pathologic findings may result from inadequate sampling that may necessitate repeat procedures (31). Thus, noninvasive neuroimaging approaches to facilitate diagnosis or to guide biopsies and surgical planning are needed to improve GBM patient care.

Through voxel-wise comparisons with histologic images, we demonstrated that DBSI-derived restricted isotropic diffusion fraction, hindered isotropic diffusion fraction, and anisotropic diffusion fraction closely correlate with high tumor cellularity, tumor necrosis, and fiber-like structures, respectively. However, these metrics alone were insufficient to clearly distinguish high tumor cellularity, tumor necrosis, or tumor infiltration (Figs. 1, 3, and 4). We, thus developed DHI, which incorporates an SVM predictive model using DBSI metrics as the classifiers, to successfully predict high tumor cellularity, tumor necrosis, and tumor infiltration against the gold standard of histology with high accuracy (Figs. 4 and 5).

Various neuroimaging techniques have been tested to assess the treatment response of brain tumors in clinical practice. Among the wide range of available neuroimaging modalities, contrast-enhanced T1WI is currently the method of choice for brain tumor diagnosis. Unfortunately, Gd-enhanced T1WI lacks specificity because it merely reflects a disrupted blood–brain barrier (32). Chemotherapy, radiation, and newer clinical trial treatments such as immunotherapies produce neuroimaging lesions that mimic tumor progression or recurrence, further confounding clinical decision-making (33). These and other shortcomings of current clinical MRI sequences suggest limitations of the MacDonald criteria (4) and the RANO updated response assessment criteria (3, 34) in monitoring tumor burden. Therefore, there is an urgent need to develop imaging modalities that can noninvasively detect and characterize the histologic features of posttreatment GBM for appropriate treatment planning.

Advanced MRI methods, such as perfusion-weighted imaging with and without contrast (35, 36) and chemical exchange saturation transfer imaging (37), and PET with amino acid tracers, including [11C]-methyl-l-methionine (38), O-(2-[18F]-fluoroethyl)-L-tyrosine (39), and 3,4-dihydroxy-6-[18F]-fluoro-L-phenylalanine (40), also provide complementary diagnostic information in GBM detection. In addition, stimulated Raman scattering microscopy (41), optical coherence tomography (42), and mass spectroscopy (43) have also been developed to improve glioma diagnosis. However, most of these techniques do not have the capability to quantify individual pathologic components noninvasively.

To address limitations of conventional MRI, diffusion-weighted MRI-derived ADC has been one of the most widely researched tools for the evaluation of tumor cellularity and grade (10). Although increased tumor cellularity has been associated with decreased ADC, the expression of aquaporin in high-grade glioma (44), vasogenic edema (45), and necrosis (46) may obviate the interpretation of expected diffusion restriction caused by high cellularity. Indeed, in our tested tumor specimens, ADC did not correlate with cellularity, while the DBSI-derived restricted fraction significantly correlated with both H&E and GFAP staining–based cellularity measures (Fig. 2). One observation in this study contradicting the widely accepted role of ADC in tumor cellularity is the significantly restricted diffusion observed in white matter tracts (Fig. 3), where the disrupted fiber network greatly increased diffusion restriction. Thus, our results further support that ADC alone cannot be considered a reliable tumor biomarker.

Through histologic validation, we demonstrated DHI accurately detects and quantifies high tumor cellularity, tumor necrosis, and tumor infiltration. The newly developed DHI framework accurately predicted key features of GBM microenvironment that eluded other neuroimaging technologies. Given the lack of specificity of clinical MRI in identifying tumor burden, DHI has the potential to aid in the noninvasive determination of tumor recurrence versus treatment response. In addition, preoperative DHI may help to guide biopsies and improve extent of resection.

J.L. Campian reports grants from NeoImmuneTech (research fund) and other from AbbVie (spouse received speaker fee from AbbVie) outside the submitted work. A.H. Kim reports grants from NIH/NINDS (R01-NS094670) during the conduct of the study, as well as grants from Monteris Medical (laser therapy grant, unrelated to this study), Stryker (dural substitute grant, unrelated to this study), and Collagen Matrix (dural substitute grant, unrelated to this study), and personal fees from Monteris Medical (consulting) outside the submitted work. S.-K. Song has ownership of Cancer Vision LLC but did not receive any types of support including grants, payments, or fees from the company. No potential conflicts of interest were disclosed by the other authors.

Z. Ye: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. R.L. Price: Conceptualization, resources, data curation, formal analysis, validation, investigation, methodology, writing-original draft, writing-review and editing. X. Liu: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. J. Lin: Data curation, formal analysis, writing-original draft. Q. Yang: Conceptualization, formal analysis, validation, writing-review and editing. P. Sun: Conceptualization, software, formal analysis, investigation, methodology, writing-review and editing. A.T. Wu: Software, formal analysis, validation, investigation. L. Wang: Software, formal analysis, investigation. R. Han: Investigation, writing-original draft, writing-review and editing. C. Song: Software, formal analysis. R. Yang: Formal analysis, funding acquisition, investigation, methodology, writing-review and editing. S.E. Gary: Formal analysis, writing-review and editing. D.D. Mao: Resources, data curation, project administration. M. Wallendorf: Software, formal analysis, writing-review and editing. J.L. Campian: Supervision, writing-review and editing. J.-S. Li: Conceptualization, supervision, writing-review and editing. S. Dahiya: Conceptualization, resources, data curation, supervision, methodology, writing-original draft, writing-review and editing. A.H. Kim: Conceptualization, resources, data curation, supervision, funding acquisition, investigation, methodology, writing-original draft, project administration, writing-review and editing. S.-K. Song: Conceptualization, resources, supervision, funding acquisition, validation, investigation, methodology, writing-original draft, project administration, writing-review and editing.

This work was supported in part by NIH (R01-NS047592, P01-NS059560, and U01-EY025500 to S.-K. Song, and R01-NS094670 to A.H. Kim), The Christopher Davidson and Knight Family Fund (to A.H. Kim), the Duesenberg Research Fund (to A.H. Kim), National Multiple Sclerosis Society (RG 1701-26617 to S.-K. Song), The Fundamental Research Funds for the Central Universities (SCUT 2018MS23 to R. Yang), Natural Science Foundation of Guangdong Province in China (2018A030313282 to R. Yang), and National Natural Science Foundation of China (81971574 to R. Yang).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.