Introduction

High-grade glioma often presents complicated tumor micro-environment containing co-existing tumor pathologies, which includes high tumor cellularity, central necrosis, micro-vascular proliferation and tumor infiltration. These pathological hallmarks set Grade IV GBM from Grade III glioma such as anaplastic astrocytoma, where the latter commonly associates with increasing tumor cellularity alone. Restricted diffusion caused by increased cellularity in glioma could be detected by DWI-derived ADC to classify glioma;¹ except for between Grades III and IV. The shared characteristic of increased cellularity between the two higher-grade gliomas requires additional spectrum for accurate diagnosis. Here we applied a novel MRI-based imaging technique called diffusion MRI histology (D-Histo), and support vector machine (SVM) to detect and classify different pathologies in glioma.

Materials and Methods

Brain tumor specimen: 14 high-grade glioma patients were recruited (Table 1) for the study. 22 brain tumor specimens were procured and fixed in 10% formalin and transferred to PBS. MRI and histological staining: Brain tumor specimens were examined using a 4.7T MR Varian scanner and a home-made surface coil. A spin-echo diffusion-weighted sequence with 99 diffusion-encoding directions with maximum b-values of 3000 s/mm² was employed to acquire DWI (in-plane resolution 0.25x0.25x0.5 mm²). The specimens underwent sequential sectioning at 5 μm thickness and was stained with H&E. Data analysis: An in-house software analyzed D-Histo and DTI data. H&E images were reviewed by an experienced neuropathologist. D-Histo maps were linearly co-registered with corresponding H&E images, allowing MRI voxels to be labeled by the histological gold standard. Machine learning: A supervised machine-learning-framework adopting SVM with RBF-kernel algorithm was used. D-Histo-derived restricted, hindered, and anisotropic fractions were used as the feature values for each voxel. We evaluated classification accuracy by pooling voxels from different classes and within each class. Receiver operating characteristics (ROC) analyses were performed to assess how well one tumor pathology could be distinguished from the rest.

Results

Figure1 displayed a representative case of a GBM specimen (Fig. 1A). T2WI, DWI and ADC map (Fig. 1B,C,D) displayed unclear or false-positive signals for dense tumor (red arrows). Hyper-intensities in D-Histo highly-restricted-fraction (Fig. 1G), restricted-fraction (Fig. 1H) and hindered-fraction (Fig. 1I) maps correspond well with tumor infiltration, dense tumor and tumor necrosis indicated by H&E and GFAP (Fig. J,K). 5220 voxels from 18 glioma specimens were distributed equally to training and validation groups to construct a robust SVM classifier. Training and validation accuracies were 93.8% and 93.9%, respectively. The 4 remaining specimens were set as testing group. D-Histo-SVM prediction (Fig. 2B) agreed highly (overall-accuracy of 95.6%) with histology classification (Fig. 2A). Prediction accuracy of the dense tumor, necrosis and tumor infiltration were 97.3%, 100% and 85.1%. Each sample was also tested to evaluate the performance of D-Histo-SVM classifier on a single subject. Representative D-Histo voxels were selected to be displayed with their corresponding H&E images. 83.6% of imaging voxels in B94 (Fig. 3A) were predicted correctly to be tumor infiltration. 98.8% and 100% of voxels in B95 (Fig. 3B) were correctly predicted to be dense tumor and necrosis. Prediction accuracies for dense tumor, necrosis and tumor in B122 (Fig. 3C) were 92.6%, 96.8% and 84.8%. 99.7% of total voxels in B128 (Fig. 3D) were correctly predicted to be dense tumor. To avoid selection bias between the training-validation and prediction groups, 7319 distinct group combinations were run, in which mean values and 95% CI were calculated (Table 2A, all samples). Dense tumor, necrosis and infiltration had average prediction accuracies of 91.7%, 93.7% and 87.8%. Although testing group was excluded when constructing the classifier, it is still possible that some of the testing data would correlate with some of the training data (i.e. they come from the same subject). We remedied this possibility by calculating 260 training-testing combinations, in which subjects in the two groups did not overlap (Table 2B, different subjects). Dense tumor, necrosis and infiltration predictions had 90.8%, 92.0% and 83.2%, accuracies. We performed ROC analyses to test D-Histo-SVM classifier’s ability to differentiate one tumor pathology from the rest (Table 2B). The classifier demonstrated AUCs of 0.978, 0.977 and 0.991 in distinguished between dense tumor and non-dense tumor, between necrosis and non-necrosis, and between infiltration and non-infiltration. Similar results were seen for different subject condition. DTI was less accurate in predicting, with lower AUC values than D-Histo (Table 2B).

Discussions and Conclusions

We demonstrated D-Histo could better reflect pathological classification of tumor grades by more accurately characterizing different tumor pathologies that are critical to clinical diagnosis and prognosis. We also differentiated tumor infiltration and tumor necrosis from other tumor pathologies, which could allow better complete tumor removal during surgery and better treatment evaluation.

Acknowledgements

This work was supported in part by NIH R01-NS047592, P01-NS059560, U01-EY025500, National Multiple Sclerosis Society (NMSS) RG 4549A4/1, RG-1507-05315.

References

1. Yamasaki F, Kurisu K, Satoh K, et al. Apparent diffusion coefficient of human brain tumors at MR imaging. Radiology. Jun 2005;235(3):985-991.