Glioblastoma (GB) is the most common and aggressive brain tumor in adults, and is characterized by a high heterogeneity within the lesion area. Lesion segmentation and classification into different tissue types, are highly important for patient diagnosis and follow up, yet is challenging using standard radiological assessment ^[1-3]. The aim of this study was to segment the enhanced and non-enhanced lesion areas and to classify them into active tumor, tissue necrosis, vasogenic edema and infiltrative disease in patients with GB, based on anatomical images and dynamic contrast enhanced MRI (DCE-MRI).

Patients and MRI Protocol: A total of 27 MRI scans obtained from fifteen patients with biopsy-proven GB were included in this study. Scans were performed on 3.0 T GE (Signa Horizon) / Siemens (Prisma) systems and included: T₁ weighted (T₁W) imaging, performed before and after contrast agent (Gd) injection; FLAIR images; and DCE, acquired using multi-phase 3D SPGR / FLASH imaging with the variable flip angle (VFA) method used for T₁ map calculation.

Preprocessing: included skull stripping, inhomogeneity correction and realigning of all anatomical images to the DCE space. T₁W+Gd and FLAIR images were normalized relative to the normal appearing white matter. Plasma volume (v_p) and the volume transfer constant (k^trens) parameters were calculated from the DCE data using DUSTER (DCE-Up-Sampled-TEmporal Resolution), an in-house tool for DCE analysis, based on the Extended-Tofts-Model ^[4] with T₁ maps calculated with correction of inaccuracy in the flip angles and accounting for differences in the bolus arrival time ^{[5, 6]}. Model selection was applied similarly to Bagher-Ebadian ^[7].

Segmentation and classification: First, the enhanced and non-enhanced lesion areas were extracted from the T₁W+Gd and FLAIR images using threshold-based segmentation similar to ^[8] (Fig. 1a,b). Next, classification of the enhanced area into active tumor or necrosis and of the non-enhanced lesion area into vasogenic edema or infiltrative disease were performed using Matlab support vector machine (SVM) classifier (Fig. 1c). The input data for classification were the normalized T₁W+Gd, FLAIR images, and the DCE v_p, and k^trens maps. Training data of active tumor (22 data sets), necrosis (8 data sets), vasogenic edema (26 data sets) and infiltrative disease (19 data sets) were manually defined retrospectively and approved by a senior neuro-radiologist, based on longitudinal MRI of the patients. Sensitivity and specificity of the training set were measured based on two-fold cross validation analysis.

Lesion segmentations and voxel-wise classification of the entire lesion area were successfully performed in all scans (n=27). Significant differences (p<0.05) were detected between the active tumor compared to necrosis with higher v_p and k^trans values detected for the active tumor, and between the vasogenic edema compared to infiltrative disease with lower FLAIR, and higher v_p and k^trans values detected for the infiltrative disease (Table1).

Validation: The classification of the SVM training data produced high sensitivities and specificities of: 97% / 94% for active tumor; 100% / 100% for necrosis; 98% / 97% for vasogenic edema and 95% / 96% for infiltrative disease. Figure 2 demonstrates data obtained from a patient with GB, scanned at time points #1 (top row), and #2 (bottom row), two months later. In this case, conventional imaging showed a mild increase of 19% of the enhanced lesion area, and a decrease of 22% of the non-enhanced lesion area, indicating stable disease according to RANO criteria ^[1]. However, the SVM segmentation results showed a marked increase of 90% in the active tumor and of 50% in the infiltrative disease components, supporting disease progression. Follow-up scans of this patients demonstrated further progression.

The current work aimed to tackle major radiological challenges of differentiating active tumor from necrosis and vasogenic edema from infiltrative disease in patients with GB. In this study automatic segmentation of lesion area and classification into four tissue types were performed using threshold based segmentation and SVM classifier, based on conventional imaging and DCE-MRI vascular parameters. In comparison to dynamic susceptibility contrast (DSC) which is usually used for assessment of brain tumors, DCE is less sensitive to susceptibility artifacts, acquired with higher spatial resolution, and provides quantitative information of different vascular properties within the tissue, including permeability. The two-stage process applied in this study showed high sensitivity and specificity, can provide volumetric information about changes in lesion components over time, and thus may be used in routine clinical practice to improve patient diagnosis and therapy response assessment.