Independent histological validation of MR-derived radio-pathomic maps of tumor cell density using image-guided biopsies in human brain tumors

Abstract

Purpose

In brain gliomas, non-invasive biomarkers reflecting tumor cellularity would be useful to guide supramarginal resections and to plan stereotactic biopsies. We aim to validate a previously-trained machine learning algorithm that generates cellularity prediction maps (CPM) from multiparametric MRI data to an independent, retrospective external cohort of gliomas undergoing image-guided biopsies, and to compare the performance of CPM and diffusion MRI apparent diffusion coefficient (ADC) in predicting cellularity.

Methods

A cohort of patients with treatment-naïve or recurrent gliomas were prospectively studied. All patients underwent pre-surgical MRI according to the standardized brain tumor imaging protocol. The surgical sampling site was planned based on image-guided biopsy targets and tissue was stained with hematoxylin–eosin for cell density count. The correlation between MRI-derived CPM values and histological cellularity, and between ADC and histological cellularity, was evaluated both assuming independent observations and accounting for non-independent observations.

Results

Sixty-six samples from twenty-seven patients were collected. Thirteen patients had treatment-naïve tumors and fourteen had recurrent lesions. CPM value accurately predicted histological cellularity in treatment-naïve patients (b = 1.4, R² = 0.2, p = 0.009, rho = 0.41, p = 0.016, RMSE = 1503 cell/mm²), but not in the recurrent sub-cohort. Similarly, ADC values showed a significant association with histological cellularity only in treatment-naive patients (b = 1.3, R² = 0.22, p = 0.007; rho = -0.37, p = 0.03), not statistically different from the CPM correlation. These findings were confirmed with statistical tests accounting for non-independent observations.

Conclusion

MRI-derived machine learning generated cellularity prediction maps (CPM) enabled a non-invasive evaluation of tumor cellularity in treatment-naïve glioma patients, although CPM did not clearly outperform ADC alone in this cohort.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11060-025-05105-x.

Introduction

High-grade gliomas, the most common primary brain malignancy [1], are primarily treated through surgical resection of the contrast-enhancing tumor, when feasible [2]. While many studies showed that a radical surgical resection of the contrast-enhancing region warrants a survival benefit [3–7], recent findings advocate for a “supramarginal” resections involving the surrounding peri-enhancing tissue, as tumor cells are known to infiltrate beyond visible enhancement and tumor often recurs in the peri-cavitary region [8–10]. When surgery is not feasible due to tumor location or multifocality, stereotactic biopsy is essential for accurate diagnosis and treatment planning [2, 11]. In both scenarios, non-invasive biomarkers of tumor cellularity could prove valuable, guiding the extent of resection in operable cases and identifying biopsy targets in inoperable ones.

Apparent Diffusion Coefficient maps (ADC) derived from diffusion-weighted imaging (DWI) is commonly considered as a non-invasive surrogate biomarker for tumor cellularity [4, 12–19], showing a moderate/strong inverse correlation with histopathological cellularity in recent meta-analyses (pooled Spearman’s correlation coefficient around –0.66) [20, 21]. ADC was proven to potentially improve the identification of areas of high cellularity in the peri-enhancing tissue [13, 22, 23]. Nevertheless, there is significant variability of findings across studies, with some authors reporting poor or even opposite relationships [24–26], possibly due to ADC values being also influenced by confounding microstructural factors other than tumor cellularity in gliomas [27].

Machine learning algorithms could represent a new approach for non-invasive tumoral cellularity prediction. Recently, a random forest ensemble algorithm was developed, that generates voxel-wise radio-pathomic maps reflecting tissue cellularity [28]. This model was trained with multiparametric conventional and diffusion MRI obtained pre-mortem to predict the histological cell count from co-registered autopsy samples from brain slices aligned with imaging. In validation studies, radio-pathomic maps could distinguish hypercellular areas within the contrast-enhancing region, identify regions of increased cellularity in the peri-enhancing tumor regions [23, 28], and provide information on who may benefit from bevacizumab [29]. Most recently, these models were applied to large publicly available imaging cohorts to explore the relationship between perfusion metrics and cell density [30].

In the current study, we aim to validate this previously-developed machine learning algorithm on an external independent cohort whose images and histopathological samples were obtained at a separate institution. In more detail, multiparametric MRI datasets served as input for the algorithm generating cellularity prediction maps (CPM), and the CPM values from image-guided biopsy targets were compared to histopathological cellularity. Moreover, since ADC is an established imaging biomarker for tumor cellularity, we compared the performance of the CPM with that of the ADC map in predicting histological cell density.

Methods

Patient selection

Patients gave written informed consent to participate in a prospective IRB-approved imaging study (IRB#14–001261). All patients had either a radiological recurrence of a previously-diagnosed glioma or a treatment-naïve radiologically suspected glioma. Pre-surgical MRI scans for these patients were acquired between April 2015 and November 2018 at the Department of Radiology of the University of California Los Angeles. All patients underwent open craniotomy for surgical resection. The precise sampling site was planned based on image-guided biopsy targets and stored for further analysis. The inclusion criteria for the current study were as follows: histopathological diagnosis of adult-type diffuse glioma, availability of pre-surgical MRI scans, and availability of digitalized hematoxylin and eosin staining (H&E).

Magnetic resonance imaging acquisition and pre-processing

All patients underwent a pre-surgical acquisition on a 3 T Siemens Prisma (Siemens Healthineers) scanner according to the standardized brain tumor imaging protocol (BTIP) [31, 32], including parameter-matched pre- and post-contrast 3D T₁-weighted images with 1-mm isotropic voxels (T1 and T1CE), 2D T₂-weighted FLAIR images with 3-mm slice thickness (FLAIR), diffusion tensor imaging (DTI) with 2-mm isotropic voxels (64 direction, b-value = 1000 s/mm²). ADC was calculated on the scanner from DTI data. A single dose of Gadavist (Gadobutrol, Bayer) (~ 0.1 mL/kg) was administered at a rate of ~ 4 mL/s during the acquisition of T₂*-weighted dynamic susceptibility images (DSC), in compliance with guidelines [33], as per BTIP [31]. Relative cerebral blood volume (rCBV) maps were calculated using a bidirectional leakage correction algorithm, as previously described [34]. DSC was not part of the machine learning algorithm.

T1, T1CE, FLAIR, and ADC images were the input sequences for the machine learning algorithm generating cellularity prediction maps (CPM), previously trained and validated on MRI and autopsy glioblastoma data [23, 28], and applied using a custom Matlab script (MathWorks). The voxel-wise CPM values output by the algorithm were divided by the whole-brain intensity SD at the single patient level, to normalize values across patients [28]. Each image-guided biopsy target consisted in a single spherical region-of-interest (ROI) with diameter 5 mm. Mean normalized CPM values and mean ADC values were extracted from each ROI.

Histopathological analysis

Samples were hematoxylin and eosin stained (H&E) and digitalized. All recurrent specimens exhibited “active” tumor cells with mitoses and ki-67 positivity, therefore no cases of pure radiation necrosis were included. QuPath (https://qupath.github.io, v05.1), an open-source software application for digital pathology, was used for semi-automatic cell counting [35]. Each stain was manually revised, and cell count parameters were adjusted to supervise the cell count. All cells identified on the H&E stain were counted and included in the cellularity quantitation, independently from their ki-67 expression.

Statistical analysis

Group comparisons. The Mann-Withey test was used to analyze group differences in CPM values (e.g., samples from contrast-enhancing tumor vs samples from non-enhancing tumor).

Cellularity correlations, assuming independent observations. The linear relationship (Pearson’s correlation) and rank correlation (Spearman’s correlation) between CPM cellularity (CPM cell/mm²) and histological cellularity in H&E stain (H&E cell/mm²) were analyzed assuming that all biopsy-derived measurements were independent. For Pearson’s correlations, observations considered as outliers were detected using ROUT method [36] and were excluded from the analysis. The performance of CPM prediction was evaluated using root mean square error (RMSE) values. Since ADC is commonly considered a proxy of tumor cellularity [24, 37], the correlations between ADC and histological cellularity were also tested. The correlation between CPM and histological cellularity was directly compared to the correlation between ADC and histological cellularity, using Meng’s z-tests. These analyses under the assumption of independent observations were conducted as a first-line approach that is easier to visualize and interpret.

Cellularity correlations, accounting for non-independent observations. Since multiple measurements were obtained from the same patient (and from the same MRI scan), the assumption of independent observations is inherently violated. To account for non-independent observations (e.g., two biopsy targets obtained from the same patient), two additional analyses were performed. First, a mixed-effects model was built to test the linear relationship between histological cellularity and CPM values (fixed effect) while accounting for the patient ID (random effect). Second, an iterative approach was performed by randomly picking a single sample per patient on each iteration, and repeated for 1000 iterations to ultimately obtain a distribution of the Spearman’s coefficients (rho) that did not include clustered observations. These analyses accounting for non-independent observations were conducted to confirm the results of the first analysis which assumed independent observations.

Results

Patients’ cohort characteristics

Twenty-seven patients met the inclusion criteria; thirteen were treatment-naïve (TN), and fourteen recurrent (R) tumors; sixty-six tumor tissue samples were collected from this cohort. Thirty-seven samples were obtained from enhancing tissue and twenty-nine from the non-enhancing tumor. Table 1 summarizes the clinical, pathological, and radiological characteristics of the enrolled cohort.

Table 1.

Demographic, clinical, and radiological characteristics

Patient characteristics	Value
Number of patients	27
Sex n (%)	M 17 (63%) F 10 (37%)
Age, (mean ± SD)	49.4 ± 15.6
Prior treatments
Treatment-naïve (%)	13 (48%)
Recurrent (%)	14 (52%)
Histology
WHO CNS4, 2016 (n, %)	Oligoastrocytoma Grade II, 1 (4%) Astrocytoma Grade II, 4 (15%) Anaplastic Oligoastrocytoma Grade III, 5 (18%) Anaplastic Oligodendroglioma Grade III, 1 (4%) Anaplastic Astrocytoma Grade III, 1 (4%) Glioblastoma Grade IV, 14 (51%) High-grade glioma Grade IV, 1 (4%)
WHO CNS5, 2021 (n, %)	Astrocytoma Grade 2, 4 (15%) Astrocytoma Grade 3, 4 (15%) Oligodendroglioma Grade 3, 1 (3%) Astrocytoma Grade 4, 5 (19%) Glioblastoma Grade. 4, 11 (41%) NOS*, 2 (7%)

Samples characteristics	Value
Number of samples	66
Location (n, %)	Frontal 27 (41%) Parietal 4 (6%) Fronto-parietal 6 (9%) Temporal 13 (20%) Occipital 5 (8%) Fronto-temporo-insular 9 (14%) Multifocal 2 (3%)
Number of samples per patient	1 sample (7 patients) 2 samples (6 patients) 3 samples (9 patients) 4 samples (5 patients)
Radiologic appearance of the sample
Enhancing (n, %)	37 (56%)
Non enhancing (n, %)	29 (44%)
H&E cellularity (cell/mm2, mean ± sd)	2856 ± 1461
CPM cellularity (cell/mm2, mean ± sd)	1323 ± 324

*In two patients it was not possible to presume the WHO CNS5 2021 classification, due to insufficient information about their molecular status

Descriptive statistics and group comparisons

The mean H&E cellularity from histology was 2856 cell/mm² (SD 1461 cell/mm²) across all surgical samples (Fig. 1a). The mean predicted cellularity value from the MRI-derived CPM map was 1323 cell/mm² (SD 324.4 cell/mm²) (Fig. 1b). Samples taken from CET regions showed significantly higher CPM values compared to the ones taken from nCET (mean: 1415 cell/mm² vs 1206 cell/mm² p = 0.007, Fig. 1c). CPM values tended to be higher in HGG than in LGG (mean: 1358 cell/mm² vs 1195 cell/mm², Fig. 1d), although this difference was not significant. Samples from IDH wild-type tumors had significantly higher CPM values compared to those from IDH mutant tumors (mean: 1432 cell/mm² vs 1243 cell/mm², p = 0.01, Fig. 1e). Finally, CPM values corresponding to samples from treatment-naïve lesions and from recurrent lesions did not exhibit a statistical difference (mean 1290 cell/mm² vs 1358 cell/mm², Fig. 1f).

Fig. 1.

Descriptive statistics of sample cellularity and group differences. A Histological cell density using H&E staining (H&E cellularity). B Cellularity predicted by the radio-pathomic maps (CPM cellularity). C–G Difference in CPM values (C) and H&E cellularity (G) between samples taken from enhancing (CET) vs non-enhancing (nCET) tumor components. D–H Difference in CPM values (D) and H&E cellularity (H) between samples taken from high- vs low-grade gliomas (HGG vs LGG). E–I Difference in CPM values (E) and H&E cellularity (I) between samples taken from IDH mutant and IDH wild-type gliomas. F–J Differences in CPM values (F) and H&E cellularity (J) between samples taken from treatment-naïve and recurrent lesions

Cellularity correlations assuming independent observations

CPM was a predictor of H&E cellularity when considering the subset of treatment-naïve lesions (TN), showing both a significant linear correlation (b = 1.4, R² = 0.2, p = 0.009, RMSE = 1503 cell/mm²) and rank correlation (rho = 0.41, p = 0.016) (Table 2, Fig. 2b). Instead, no significant association was found between CPM values and H&E cellularity in the whole cohort (b = 0.7, R² = 0.028, p > 0.05, RMSE 1974 cell/mm²; rho = 0.17, p > 0.05, Fig. 2a), and considering only patients with recurrent disease (b = -0.66, R² = 0.014, p > 0.05, RMSE 3156 cell/mm²; rho = -0.12, p > 0.05, Fig. 2c).

Table 2.

Comparison between radio-pathomic map and ADC statistics

	Metrics	Value (p)
		Whole cohort	Treatment-naïve	Recurrent
CPM	b (Pearson’s)	0.7 (> 0.05)	1.4 (0.009)*	− 0.66 (> 0.05)
	R2 (Pearson’s)	0.028	0.20	0.014
	rho (Spearman’s)	0.17 (> 0.05)	0.41 (0.016)*	− 0.12 (> 0.05)
	RMSE (cell/mm2)	1974	1503	3156
ADC	b (Pearson’s)	− 0.78 (> 0.05)	− 1.3 (0.007)*	0.78 (> 0.05)
	R2 (Pearson’s)	0.032	0.22	0.031
	rho (Spearman’s)	− 0.18 (> 0.05)	− 0.37 (0.03)*	− 0.18 (> 0.05)

b regression coefficient, R² coefficient of determination, Rho Spearman’s correlation coefficient, RSME root mean square error, ns not significant

Fig. 2.

Relationship between CPM cellularity from radio-pathomic maps, ADC values, and histological H&E cellularity. The radio-pathomic map performs well in predicting the histological cellularity of treatment-naïve tumors (B), while no association was found in recurrent patients (C) and in the whole cohort (A). Similarly, ADC was associated with histological cellularity only in the treatment-naïve cohort (E), and not in the whole cohort (D) nor the recurrent subset (F). White dots represent outliers

Similarly, the ADC map showed a significant association with H&E cellularity in treatment-naïve patients (b = 1.3, R² = 0.22, p = 0.007; rho = -0.37, p = 0.03, Fig. 2e), while no association was found in the whole cohort (b = -0.78, R² = 0.032, p = 0.15; rho = -0.18, p > 0.05, Fig. 2d) and in recurrent lesions (b = 0.78, R² = 0.031, p > 0.05; rho = -0.18, p > 0.05, Fig. 2f).

The correlation between CPM and histological cellularity was not statistically different from the correlation between ADC and histological cellularity (z = 0.27, p = 0.79 in the treatment-naïve subgroup).

Cellularity correlations accounting for non-independent observations

Additional analyses accounting for non-independent observations confirmed the results of the independent sample approach. With the linear mixed model (Supplemental Table 1), CPM values were statistically-significant predictors of H&E cellularity only in the treatment-naïve cohort (b = 1.41, p = 0.013), and not significant predictors in the whole cohort and in recurrent cases. Similarly, the association between ADC values and H&E cellularity was significant only in treatment-naïve patients (b = -1.21, p = 0.01).

The results of the iterative approach (Supplemental Fig. 1) revealed a higher correlation between CPM and H&E cellularity in treatment-naïve patients (mean rho = 0.51) than in the entire cohort (mean rho = 0.25) or in the recurrent group (mean rho = 0.22). Similarly, the association between ADC values and H&E cellularity was stronger in treatment-naïve patients (mean rho = -0.37) compared to the entire cohort (mean rho = -0.22) and to the recurrent group (mean rho = -0.08).

Representative cases

A 86-year-old female received a first diagnosis of right temporal lobe glioblastoma (IDH wild-type, grade 4, Fig. 3a). In the targeted surgical region, obtained from contrast-enhancing tumor tissue, CPM maps exhibited relatively low values of predicted cellular density (916 cells/mm²), and ADC was 1145 × 10^–6 mm²/s. Histological analysis revealed a relatively low tumor cellularity (2112 cells/mm²). A 28-year-old male had a treatment-naïve non-enhancing astrocytoma in the left fronto-parietal area (IDH mutated, 1p19q intact, grade 3, Fig. 3b). In the targeted region, CPM maps exhibited high values of predicted cellular density (1983 cells/mm²), and ADC was 1127 × 10^–6 mm²/s. Histological analysis revealed a relatively high tumor cellularity (4749 cells/mm²). These two cases show successful prediction of H&E histology using CPM maps.

Fig. 3.

Representative cases of good agreement between radio-pathomic maps and histology. A Sample collected from the enhancing tumor component. CPM values were low, and this finding was associated with low histological cellularity at H&E stain. Interestingly rCBV values were not elevated. B Sample taken from non-enhancing tumoral tissue. CPM values are elevated, and a high histological cellularity is found on H&E. rCBV values were increased in this region

A 35-year-old female presented with a treatment-naïve left frontal glioblastoma (IDH wild-type, grade 4, Fig. 4a). In the targeted surgical sample, obtained non-enhancing tumor tissue, CPM maps exhibited moderate values of predicted cellular density (1240 cells/mm²) and ADC was 685 × 10^–6 mm²/s. However, histological analysis revealed extremely high tumor cellularity (7449 cells/mm²). A 24-year-old female received a first diagnosis of right fronto-temporo-insular astrocytoma (IDH mutated, 1p19q intact, grade 4, Fig. 4b), non-enhancing. In the targeted region, CPM maps exhibited moderate values of predicted cellular density (1354 cells/mm²) and ADC was 1992 × 10^–6 mm²/s. However, histological samples revealed a high tumor cellularity (5732 cells/mm²). In these latter two cases CPM map fail to accurately predict histological cellularity.

Fig. 4.

Representative cases of poor agreement between radio-pathomic maps and histology. A Sample collected from non-enhancing tumoral tissue. In this case, the CPM map fails to predict the high cellularity found in histological analysis. B Sample collected from non-enhancing tumoral tissue. The CPM map shows low values of predicted cellularity, while H&E showed a moderate to high histological cellularity. In both cases, rCBV values were not elevated

Discussion

Results from this study suggest that CPM values provide valuable information about tumor cellularity in treatment-naïve patients, while the recurrent sub-cohort showed poor association with histological findings. Similarly, ADC values showed an association with histological tumor cellularity in the treatment-naïve sub-cohort (with a slightly weaker correlation compared to CPM), but not in the recurrent sub-cohort.

In treatment-naïve gliomas, identifying highly cellular areas, particularly beyond the contrast-enhancing margin, could help to depict the areas of tumoral infiltration, guiding surgical resection and optimizing adjuvant treatment, as radiotherapy [2]. Maximal safe resection improves patients’ prognosis [2], and recent evidence demonstrates that the additional removal of infiltrating tumor components in the peri-enhancing region (“supramarginal resection”), whenever feasible, leads to a better prognosis in high-grade gliomas [38–41]. Since peri-enhancing T2-weighted/FLAIR abnormalities alone do not directly reflect tumor cellularity [38, 41, 42], imaging biomarkers depicting areas of increased cellularity in the peri-enhancing region would enable surgeons to remove infiltrated tumor tissue while minimizing the risk of postoperative neurological deficits related to the increased extent of resection. Interestingly, 34% of samples in our study obtained from high-grade gliomas were collected from non-enhancing areas. The RMSE calculated only in this subset of samples (1777 cells/mm²) was lower than the ones of the whole cohort (1974 cell/mm²) indicating a smaller error of the model in the prediction of the histopathologic cell density in this specific group (Supplemental Fig. 2). This finding is particularly noteworthy, suggesting that CPM could serve as a valuable biomarker for predicting tumor cellularity beyond the contrast-enhancing regions, potentially guiding “supramarginal” resections. In addition, non-invasive biomarkers reflecting tumor cellularity would also be useful in non-resectable treatment-naïve tumors, as they would improve the planning of stereotactic biopsies, which should target the most aggressive areas for a more accurate histopathological diagnosis.

ADC from DWI is commonly used as a non-invasive biomarker for tumor cellularity, but its correlation with histological cellularity is highly variable in the literature [20, 21, 24–26, 43]. This could be due to microstructural factors other than cell density, that influence ADC values, such as extracellular matrix composition, vasogenic edema, degenerative changes, and tissue compression [27, 44, 45]. CPM demonstrated a positive moderate correlation with H&E cellularity in treatment-naïve patients, possibly outperforming ADC when accounting for non-independent observations.

In the recurrent setting, too, the non-invasive evaluation of tumor cellularity is clinically relevant. After adjuvant chemoradiation, distinguishing between true tumor progression and treatment-induced effects (also referred to as “pseudoprogression”, and which can include radiation necrosis) is challenging [46]. Current solutions include advanced imaging such as perfusion-weighted imaging and amino-acid PET, and the use of confirmatory scans as suggested in the Response Assessment in Neuro-Oncology (RANO) criteria 2.0 [47, 48]. However, confirmation scans delay the diagnosis of tumor progression [49], and the routine use of advanced imaging is limited to a few centers and suffers from technical limitations that limit their universalizability [50]. Since true tumor progression is marked by actively mitotic tumor cells, linked to an increase in local cellularity [51–54], imaging surrogates of cellularity would arguably aid the distinction between tumor progression and pseudoprogression. Some meta-analyses reported encouraging results supporting the role of diffusion imaging for this application [55, 56], but a pooled cut-off ADC value obtained from multiple studies (1330 × 10^–6 mm²/s) showed a poor diagnostic performance overall, due to an overlap of ADC values between groups [56]. CPM maps could represent a valid alternative to evaluate tumor cellularity and identify areas of true tumor progression when monitoring treated gliomas. However, CPM performed worse in recurrent patients, and the RMSE value in this subset (3156 cell/mm²) was remarkably higher than in the treatment-naïve subset (1503 cell/mm²), indicating a greater error of the model in the prediction of histopathological cellularity in this specific subgroup. This increased error, as well as a non-significant correlation, shows that CPM maps may not be reliable predictors of histopathological cellularity in the recurrent setting. This finding can appear counterintuitive, since the CPM machine learning algorithm was trained on datasets collected mostly from recurrent tumor [28]. However, the scenario of chemoradiation-treated recurrent glioblastomas is complex from a radiographic standpoint, since actively-growing tumor tissue is often mixed with areas of treatment-related effects, presenting overlapping MRI features. Additionally, our patients with recurrent lesions received different therapies, which potentially induced heterogenous radiographic changes. Overall, these factors related to prior treatments may explain why the current CPM model may not reliably predict cellularity in this patient subset. Notably, in line with prior meta-analyses [55, 56], ADC was not a significant predictor of tumor cellularity in the recurrent sub-cohort in our study, either. Our findings, as well as data from the literature, suggest that currently MRI-based prediction of tumor cellularity in the recurrent setting with either ADC or CPM may be unsatisfactory. Integrating multi-parametric imaging, such as perfusion-weighted imaging, metabolic imaging, or advanced diffusion models, into new machine learning algorithms predicting tumor cellularity may potentially help overcoming this limitation, by providing more pathophysiological information related to active tumor growth and treatment effects.

Limitations

The relatively small sample size of our cohort could have reduced the power of the statistical analyses. The diverse histological characteristics within this cohort may have introduced additional variability based on the different tumor subtypes. The samples were collected during surgery using pre-operative imaging. However, some inherent limitations in the precision of the correspondence between pre-operatory imaging and sampling site were inevitable, due to the changes in the operative field related to brain shift and surgical manipulation.

Conclusion

In this study, we tested a previously-developed machine learning algorithm for MRI-based cellularity prediction on an external independent cohort and compared its performance with that of the ADC maps. Machine learning generated cellularity prediction maps (CPM) were valuable for tumor cellular prediction in treatment-naïve patients in samples collected from both contrast-enhancing and non-enhancing areas. In recurrent patients, the association between CPM and histological cellularity was poor. Our results suggest that machine learning algorithms exploiting the information from multi-parametric MRI data have the potential to enhance tumor cellularity prediction in gliomas, with potential applications in guiding stereotactic biopsy sampling and identifying tumor-infiltrated non-enhancing areas for tailored supramarginal surgical resections. However, the performance of CPM maps did not show clear superiority over the ADC maps for the prediction of histological cellularity in this cohort.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Study design: GN, FS, BME. Data collection: GN, FS, BME, JY, CR, KP, RGE, LML, NS, TFC. Data curation: GN, FS. Data processing: GN, FS, BME, JY, SAB, JC, PSLV. Visualization: FS, GN, BME. Statistical analysis: FS, GN, BME. Interpretation: all authors. Manuscript initial draft: GN, FS, BME. Manuscript revision and editing: all authors.

Funding

This study was supported by grant from the National Brain Tumor Society (NBTS) (Ellingson, Cloughesy), the Sontag Foundation (Ellingson, Cloughesy), NIH/NCI R01CA270027 (Ellingson, Cloughesy), NIH/NCI R01CA279984 (Ellingson, Cloughesy), DoD CDMRP CA220732 (Ellingson, Cloughesy), NIH/NCI P50CA211015 (Ellingson, Cloughesy), Nakatini Foundation (Oshima), Strain for the Brain (Bobholz, LaViolette), R01CA218144 (LaViolette, Connelly, Bobholz), R01CA249882 (Bobholz, LaViolette), R01CA290631 (LaViolette, Connelly, and Bobholz), Advancing a Healthier Wisconsin (LaViolette), American Brain Tumor Association (LaViolette), and the Ryan M. Schaller Foundation (LaViolette, Bobholz).

Data availability

Data from this cohort is available from the authors upon request.

Declarations

Competing interests

BME is on the advisory board and is a paid consultant for Alpheus Medical, Carthera, Chimerix, Ellipses Pharma, Erasca, Global Coalition for Adaptive Research (GCAR), Imaging Endpoints, Medicenna, Voiant, Medscape, Monteris, Neosoma, Nerviano Medical Sciences, Nuvation Bio, Orbus Therapeutics, Sagimet Biosciences, Sapience Therapeutics, Servier Pharmaceuticals, Siemens, SonALAsense, Sumitomo Dianippon Pharma Oncology, Telix, the Sontag Foundation, the National Brain Tumor Society, and Third Rock Ventures. TFC is cofounder, major stock holder, consultant and board member of Katmai Pharmaceuticals, holds stock for Erasca, member of the board and paid consultant for the 501c3 Global Coalition for Adaptive Research, holds stock in Chimerix and receives milestone payments and possible future royalties, member of the scientific advisory board for Break Through Cancer, member of the scientific advisory board for Cure Brain Cancer Foundation, has provided paid consulting services to Blue Rock, Vida Ventures, Lista Therapeutics, Stemline, Novartis, Roche, Sonalasense, Sagimet, Clinical Care Options, Ideology Health, Servier, Jubilant, Immvira, Gan & Lee, BrainStorm, Katmai, Sapience, Inovio, Vigeo Therapeutics, DNATrix, Tyme, SDP, Kintara, Bayer, Merck, Boehinger Ingelheim, VBL, Amgen, Kiyatec, Odonate Thera- peutics QED, Medefield, Pascal Biosciences, Bayer, Tocagen, Karyo- pharm, GW Pharma, Abbvie, VBI, Deciphera, VBL, Agios, Genocea, Celgene, Puma, Lilly, BMS, Cortice, Novocure, Novogen, Boston Biomedical, Sunovion, Insys, Pfizer, Notable labs, Medqia, Trizel, Medscape and has contracts with UCLA for the Brain Tumor Program with Roche, VBI, Merck, Novartis, BMS, AstraZeneca, Servier. The Regents of the University of California (T.F.C. employer) has licensed intellectual property co-invented by TFC to Katmai Pharmaceuticals. PSL Holds US Patent 12171542 that protects portions of the intellectual property of the radio-pathomic mapping methods used in this study.

Ethical approval

The collection and analysis of clinical and imaging data for this research was approved by the institutional review board under the identification number IRB#14-001261.

Informed consent

All patients gave written informed consent for their datasets and clinical information to be included in a database used for research.