Visualization of heterogeneity and regional grading of gliomas by multiple features using magnetic resonance-based clustered images

Preoperative glioma grading is important for therapeutic strategies and influences prognosis. Intratumoral heterogeneity can cause an underestimation of grading because of the sampling error

in biopsies. We developed a voxel-based unsupervised clustering method with multiple magnetic resonance imaging (MRI)-derived features using a self-organizing map followed by K-means. This

method produced novel magnetic resonance-based clustered images (MRcIs) that enabled the visualization of glioma grades in 36 patients. The 12-class MRcIs revealed the highest classification

performance for the prediction of glioma grading (area under the receiver operating characteristic curve = 0.928; 95% confidential interval = 0.920–0.936). Furthermore, we also created

12-class MRcIs in four new patients using the previous data from the 36 patients as training data and obtained tissue sections of the classes 11 and 12, which were significantly higher in

high-grade gliomas (HGGs) and those of classes 4, 5 and 9, which were not significantly different between HGGs and low-grade gliomas (LGGs), according to a MRcI-based navigational system.

The tissues of classes 11 and 12 showed features of malignant glioma, whereas those of classes 4, 5 and 9 showed LGGs without anaplastic features. These results suggest that the proposed

voxel-based clustering method provides new insights into preoperative regional glioma grading.

Gliomas are the most common primary brain tumor with poor prognosis and are graded according to the classification by the World Health Organization (WHO) of the most malignant region. Tumor

grading is important for deciding the treatment, including surgical resection, adjuvant radiation and chemotherapy. The 5-year survival rate of patients with low-grade gliomas (LGGs) (grade

II) is 42–92%1, whereas patients with high-grade gliomas (HGGs) (grades III and IV) have a worse prognosis2. In particular, glioblastomas (grade IV) develop rapidly3 and the 5-year survival

rate of patients with these tumors is only 2%4. The histologic findings of glioblastomas include diffuse infiltration and simultaneous necrosis in different parts of the tumor. Owing to

their heterogeneity, the initial diagnosis by biopsy differs from the diagnosis by total resection in 38% of cases5. Previous studies have attempted to grade tumors as a whole; therefore,

they could not obtain local information on tumors. If the grade of each region is identified preoperatively, then neurosurgeons can clarify the target for biopsy and the parts of the tumor

that need to be resected or preserved, thereby preserving motor or language function.

Magnetic resonance imaging (MRI) is essential for noninvasive diagnosis of the existence, extent and characteristics of brain tumors. T1-weighted imaging (T1WI), contrast-enhanced

T1-weighted imaging (T1WIce), T2-weighted imaging (T2WI), fluid-attenuated inversion recovery (FLAIR) and diffusion-weighted imaging (DWI) sequences are generally performed before surgery.

These conventional images can yield a large amount of useful information on tumors, such as tumor morphology, the presence of enhancement, intratumoral hemorrhage, or edema and are helpful

for predicting the tumor grade. It is still controversial whether the presence of enhancement indicates malignancy6,7,8. Contrast enhancement is nonspecific to malignancy and primarily

reflects the passage of contrast material across a disrupted blood–brain barrier9. Despite this uncertainty, we often examine T1-enhancing regions to determine the target area for the

biopsy. Therefore, the most malignant tissue section may not be obtained. The peritumoral hyperintensity on T2WI or FLAIR is also nonspecific, representing tumor infiltration, vasogenic

cerebral edema, or both10,11. These previous studies suggest that glioma grading using only a single feature in MRI is not accurate.

Recently, some studies have applied pattern recognition methods using multiple features to predict tumor grading12,13. However, these methods demonstrate several problems in clinical

application. Pattern recognition methods have two types of clustering, namely supervised and unsupervised and previous studies have applied either of them. Supervised methods require a

priori knowledge of boundaries and tissue signatures, whereas unsupervised methods do not14. Because methods for preoperative assessment of tumor pathology have not yet been developed,

supervised labeling can be inaccurate. Although unsupervised methods are useful and not arbitrary, complicated features, such as those used in previous studies12,13, make it difficult for

clinicians to recognize the most sensitive features for grading. Therefore, unsupervised techniques with multiple and simple features are more useful and suitable for clinical application.

We have recently reported that the multiple features calculated from voxel-based diffusion tensor images (DTI) and clustered by a two-level clustering approach, an unsupervised clustering

method with a self-organizing map (SOM)15 followed by K-means (KM)16, can effectively differentiate between LGGs and HGGs17. SOM is a well-known type of neural network unsupervised

learning15 that simplifies features and shows good visualization of results for data understanding and survey using component planes18,19. In addition, features that have similar patterns

can be identified by KM clustering for the results of SOM. This two-level clustering approach has two important benefits in terms of noise reduction and computational cost. First, because

the KM algorithm is very sensitive to outliers20, any outliers can adversely affect the accuracy of the clustering. When an SOM is applied prior to KM, outliers can be filtered out,

improving the clustering accuracy. Second, the computational time of the two-level clustering approach is considerably shorter than that of KM alone17. However, DTI has some limitations such

as distortions induced by susceptibility artifacts and low-spatial resolution; hence, the use of DTI in clinical situations or in a navigational system becomes difficult, thereby

automatically providing real-time cross-sectional images on conventional MRIs of the brain intraoperatively. To circumvent the problems of DTI, imaging methods that show less distortion and

higher spatial resolution, such as T1WI, T1WIce, T2WI and FLAIR, were used in this study. The higher spatial resolution of these images is suitable for an intraoperative neuronavigational

system because the preoperative voxel-based clustered images will have been correctly combined with the preoperative conventional MRI as much as possible to obtain tissue section for each

class. Furthermore, the high-spatial resolution leads to the possibility of regional glioma grading. We aimed to develop novel MR-based clustered images (MRcIs) using our two-level

clustering approach with multiple features in conventional MRI, such as T1WI, T1WIce, T2WI and FLAIR and to use those images to visualize the regional glioma grading because these

conventional MRIs are more familiar to neurosurgeons and are easier to understand. We also aimed to determine if our method can correctly predict glioma grading preoperatively in a

supervised manner. We demonstrate the possibility of regional glioma grading confirmed by pathological examination with an MRcIs-based neuronavigational system in new patients with glioma.

We retrospectively reviewed 36 patients, including 21 patients with HGGs and 15 patients with LGGs (Table 1), who underwent T1WI, T1WIce, T2WI and FLAIR before tumor resection. An overview

of the study procedure is depicted in Fig. 1. The component planes of the four MRI variables from T1WI, T1WIce, T2WI and FLAIR by the SOM analysis show the information of each sequence in

each map unit as well as the associations between the clusters and variables18 (Fig. 2). The component planes differed from each other. For example, for 12 clusters, T2WI values in the

classes 1–5 and 9 were higher than those of the other classes, whereas FLAIR values in the classes 4 and 9 were higher than those of the other classes. T1WIce values in the classes 11 and 12

and T1WI values in class 7 were higher than those of the other classes.

Component planes with SOM ranging from blue to red according to intensities in each MRI value.

The inter-class borderlines obtained by KM++ with K = 12 are shown on the SOM component planes as white lines between the nodes. Detailed intensity profiles can be seen on the SOM component

planes and patterns in each class (from 1 to 12) on the illustrative map (lower right).

Figure 3 shows representative cases of LGGs and HGGs. Although the voxels of the classes 11 and 12 can be seen in the abnormal areas of LGGs and HGGs, most of them were linearly consecutive

and the raw T1WIce showed that some of them were enhanced vessels. However, in HGGs, voxels of the classes 11 and 12 without vessels were observed, which spread like a stain within the tumor

(pink arrow in Fig. 3). Conversely, enhanced regions were not always seen to cluster in the classes 11 or 12, particularly in LGGs (light blue arrow in Fig. 3). Thus, a clear

differentiation between LGGs and HGGs could be observed on the MRcIs.

Representative cases of low- and high-grade gliomas, including the 12-class MRcIs, which showed the highest classification performance.

The MRcIs, T1WI, T1WIce, T2WI and FLAIR are shown for each patient. Each color on the MRcIs corresponds to each class in the 12-color bar (left hand corner). Inside the enhanced tumor

regions, Classes H and L/N are shown in red and green, respectively, on the enlarged T1WIce (right).

The results of leave-one-out cross-validation (LOOCV) that was used to assess the classification performances using MRcIs and support vector machines (SVM) are shown in Fig. 4A (left). The

differences in the areas under the curve (AUCs) were significant among the classes [F(6, 693) = 1147.4; p