Multi-parametric and multi-regional histogram analysis of MRI: modality integration reveals imaging phenotypes of glioblastoma

From July 2010 to August 2015, suspected patients with supratentorial newly diagnosed glioblastoma were prospectively recruited. Patients were required to have good performance status (World Health Organization performance status 0-1) before surgery. Patients were excluded when they had the history of previous brain tumor, cranial surgery, radiotherapy/chemotherapy, or contraindication for MRI scan. This study was approved by the local institutional review board. Signed informed consent was obtained from all patients.

Neuronavigation and 5-aminolevulinic acid fluorescence were used to guide surgery, with other adjuvants (e.g., cortical and subcortical mapping, awake surgery, and intraoperative electrophysiology, when appropriate) to allow maximal safe resection. Extent of resection was assessed according to postoperative MRI within 72 h, classified as complete or partial resection of enhancing tumor or open biopsy [15]. Patient postoperative treatment was determined by the multi-disciplinary team in each case according to their postoperative status. All clinical, radiological, and histological data were collected prospectively. Clinical and radiological data were incorporated to evaluate response according to the Response Assessment in Neuro-oncology criteria [2]. Specifically, within the first 12 weeks after the completion of radiotherapy, progression was only determined if new enhancement was predominantly outside of the radiation field, unless pathological evidence was available. When patients were suspected of having pseudoprogression, our multi-disciplinary team continued current treatment, with close observation, until new evidence of true progression was confirmed. Therefore, for these patients, the progression-free survival was determined retrospectively.

MRI sequences were acquired on a 3-Tesla scanner (Magnetron Trio; Siemens Healthcare, Erlangen, Germany) with a standard 12-channel receive-head coil. MRI sequences included post-contrast T1W, T2W, T2W-FLAIR, DSC, DTI, and multi-voxel 2D ¹H-MRS chemical shift imaging (CSI). PRESS excitation was selected to encompass a grid of 8 rows × 8 columns on T2W images. The scanning details are in Supplementary Methods.

All images were co-registered to T2W images, where the CSI was planned. The image co-registration was performed using the linear registration tool (FLIRT) in Oxford Centre for Functional MRI of the Brain Software Library (FSL) v5.0.0 (Oxford, UK) [16, 17]. DSC was processed and leakage correction was performed using NordicICE (NordicNeuroLab). The arterial input function was automatically defined. The rCBV, MTT, and rCBF maps were calculated. DTI was processed using the diffusion toolbox in FSL [18]. Normalization and eddy current correction were performed. The isotropic (p) and anisotropic (q) components were calculated [6].

Tumor ROIs were manually delineated on the post-contrast T1W and FLAIR images using 3D slicer v4.6.2 (https://​www.​slicer.​org/​) [19]. The delineation was independently performed by a neurosurgeon (CL, > 8 years of experience), and a researcher (NRB, > 4 years of image analysis experience), and reviewed by a neuroradiologist (TM, > 8 years of experience). Each author used consistent criteria in each patient, and was blinded to the patient clustering and outcomes. Non-enhancing (NE) ROIs were defined as the non-enhancing regions outside of contrast-enhancing (CE) regions, obtained by a Boolean subtraction of CE and FLAIR ROIs in MATLAB (version 2016a, MathWorks, Inc.). For each subject, regions of normal-appearing white matter were manually segmented in contralateral white matter and used as normal controls [20]. This region was typically located in the white matter which was furthest in distance from the tumor location and has no perceivable abnormalities. Each tumor voxel was normalized by dividing it by the mean value of normal-appearing white matter. Inter-rater reliability was performed using Dice similarity coefficient scores.

The study design is summarized in Fig. 1. Histogram features were extracted using MATLAB. Perfusion images (rCBV, MTT, and rCBF) and diffusion images (DTI-p and DTI-q) were analyzed separately. The CE and NE ROIs were also analyzed independently. Therefore, four categories of feature sets (CE-diffusion, NE-diffusion, CE-perfusion, NE-perfusion) were obtained. Intensity histograms were constructed using 100 bins. As shown in Fig. 1, a total of 10 histogram features were calculated, including mean, standard deviation (SD), median, mode, skewness, kurtosis, and 5th (Prc5), 25th (Prc25), 75th (Prc75), and 95th (Prc95) percentiles. Therefore, altogether 100 features were extracted from each subject.

The analysis was performed using a multi-view late integration methodology called multi-view biological data analysis (MVDA), implemented in R and available from GitHub (https://​github.​com/​angy89/​MVDA). Late integration methodologies allow analyzing each view independently and then merging the results [21].

The analysis was divided into multiple steps [12]. Briefly, the features were firstly clustered using the hierarchical ward clustering method in each view. The number of features was reduced by selecting the centroids of feature clusters, according to the correlation within each feature cluster. Next, for each view, the patients were clustered by applying a hierarchical ward clustering method using the features selected from last step. Lastly, the clustering results of each view were integrated in a late integration method to yield two final patient clusters. Analysis details are in Supplementary Methods.

To validate patient clustering was not obtained by random, a leave-one-out cross-validation (LOOCV) procedure was applied. Briefly, all steps of MVDA approach were repeated by leaving one patient out at each repetition. The consensus analysis was performed in the 80 clustering results obtained from the LOOCV approach. An 80 × 80 co-occurrence consensus clustering matrix M was created, where M (i, j) indicating percentage of times that the patients i and j were clustered together across the 80 dataset perturbations.

CSI data were processed using LCModel (Provencher). Choline (Cho) and N-acetylaspartate/creatine (NAA) were calculated as a ratio to creatine (Cr). All relevant spectra from CSI voxels were assessed for artifacts using described criteria [22]. The values of the Cramer-Rao lower bounds were used to evaluate the quality and reliability of CSI data. CSI values with SD > 20% were discarded. The DTI and PWI images were co-registered to T2-space, which was used to plan CSI acquisition. To account for the resolution difference between T2- and CSI space, all co-registered data were projected to CSI space, according to their coordinates using MATLAB. The proportion of T2-space tumor pixels occupying each CSI voxel was calculated. Only CSI voxels containing more than 50% tumor T2-voxels were included for further analysis.

To estimate the contribution of each centroid feature in the clustering, the variable importance evaluation function “varImp” in the R package “Caret” was used [23]. The patient clustering result was firstly used to train a logistic regression model, which was used to evaluate the importance of each feature, according to the model performance. The feature importance was scored and scaled with a maximum value of 100.

All analyses were performed in RStudio v3.2.3. CSI data were compared with Wilcoxon rank sum test using Benjamini-Hochberg procedure for controlling the false discovery rate in multiple comparisons. Kaplan-Meier and Cox proportional hazards regression analyses were performed to evaluate patient survival. For Cox proportional hazards regression, all relevant covariates, including IDH-1 mutation and MGMT methylation status, sex, age, extent of resection, and contrast-enhancing volume were considered. For Kaplan-Meier analysis using log-rank test, each feature was dichotomized using optimal cutoff values calculated by “surv_cutpoint” function in the R Package “survminer.” Patients alive at the last known follow-up were censored. Logistic regression was used to test prognostic values of covariates for 12-month overall survival (OS) and progression-free survival (PFS). The baseline models were constructed using all relevant clinical covariates. Histogram features were subsequently added into baseline models. The incremental prognostic values of the models with histogram features were determined by comparing AUC using one-way ANOVA. The hypothesis was accepted at a two-sided significance level of alpha = 0.05.

A total of 136 patients were recruited for MRI scan and surgery. Twenty-one patients were excluded due to non-glioblastoma pathology diagnosis after surgery. Postoperatively, patients received concurrent temozolomide chemoradiotherapy followed by adjuvant temozolomide following the Stupp protocol (73.0%, 84), short-course radiotherapy (17.4%, 20/115), or best supportive care (9.6%, 11/115), respectively. Among the 84 patients, four patients were lost in follow-up and excluded. A total of 80 patients were finally included into the study. Characteristics of 80 patients and two patient clusters were summarized in Table 1.

Inter-rater reliability testing of ROIs showed fair agreement between the raters, with Dice scores 0.85 ± 0.10 (mean ± SD) for CE and 0.86 ± 0.10 of FLAIR ROIs respectively.

From the four views, 5, 4, 7, and 6 centroid features were respectively selected (Fig. 2, Table 2). Using the centroid features and optimal cluster numbers determined in the algorithm, patients were firstly divided into 7, 8, 9, and 10 clusters in each view, using the hierarchical ward clustering. Late integration of four views yielded a final clustering of two patient clusters, with 53 and 27 patients in each cluster respectively.

After leave-one-out cross validation, the co-occurrence consensus clustering matrix was computed. The result showed that the two patient clusters generated from the unsupervised clustering were stable. The mean values of the co-occurrence consensus clustering matrix were 0.79 for Cluster 1 and 0.68 for Cluster 2 (Fig. 3).

The two patient clusters showed no significant differences in clinical characteristics (Table 1). Interestingly, two clusters had similar contrast-enhancing tumor volume (p = 0.823). Cluster 1, however, had significantly smaller non-enhancing tumor volume (p = 0.007) than Cluster 2. Further, the two clusters showed significant difference in survival. Specifically, Cluster 2 showed better OS (log-rank test, p = 0.020) and better PFS (log-rank test, p < 0.001) than Cluster 1 in Kaplan-Meier analysis (Table 1, Fig. 4a and b).

Since MGMT promoter methylation status was missing in two patients, the multi-variate Cox proportional hazards regression modeling was tested in the remaining 78 patients. The results showed that Cluster 2 displayed significantly better OS (p = 0.007, HR = 0.32) and PFS (p < 0.001, HR = 0.33) than Cluster 1, considering relevant covariates. Among these covariates, extent of resection (p = 0.019, HR = 2.20) and contrast-enhancing tumor volume (p < 0.001, HR = 1.02) significantly affected OS. Extent of resection (p = 0.003, HR = 2.84) significantly affected PFS. No significance was found in other clinical factors.

Due to the abovementioned rules excluding CSI voxels containing less than 50% tumor, CSI data were missing in four patients. Our results showed NAA/Cr ratio in NE region of Cluster 2 was significantly lower than Cluster 1 (p = 0.040) after controlling multiple comparisons. The comparison of CSI data in two patient clusters are detailed in Supplementary Table S1.

Seven features with a score over 50 were selected, according to the importance of centroid features in the clustering (Fig. 5). All the seven features showed significance in survival analysis (Table 3). Particularly, higher Mean-q-NE was associated with worse OS (HR = 1.40, p = 0.020) and worse PFS (HR = 1.36, p = 0.031). The Kaplan-Meier curves showing the relevance of Mean-q-NE in OS and PFS are demonstrated by Fig. 4c and d.

For prediction of 12-month OS and PFS, the AUC of baseline multi-variate models were 0.81 (95% CI, 0.70–0.93) and 0.77 (95% CI, 0.65–0.88) respectively. The results of model comparison showed that the seven features significantly improved both OS model (AUC, 0.91; 95% CI, 0.84–0.99; p = 0.020) and PFS model (AUC, 0.89; 95% CI, 0.81–0.97; p = 0.022) (Fig. 6).