Glioblastoma and primary central nervous system lymphoma: Preoperative differentiation by using MRI-based 3D texture analysis

https://doi.org/10.1016/j.clineuro.2018.08.004Get rights and content

Highlights

  • We evaluated the utility of 3D texture analysis in discrimination of GBM and PCNSL.

  • The shape features showed no differences between GBMs and PCNSLs.

  • The Firstorder_Skewness was the best selected predictor for classification.

  • The diagnostic accuracy and AUCs varied from 0.82 to 0.88, 0.85 to 0.90, respectively.

Abstract

Objectives

To investigate the diagnostic value of magnetic resonance imaging (MRI)-based 3D texture and shape features in the differentiation of glioblastoma (GBM) and primary central nervous system lymphoma (PCNSL).

Patients and methods

A total of eighty-two patients, including sixty patients with GBM and twenty-two patients with PCNSL were followed up retrospectively from January 2012 to September 2017. MRI-based 3D texture and shape analysis were performed to evaluate the detectable differences between the two malignancies. The performance of machine-learning models was assessed. The Mann-Whitney U test and receiver operating characteristic (ROC) analysis were performed, and the corresponding sensitivity, specificity, accuracy, and area under the curve (AUC) were calculated.

Ultimately, 60 GBM patients (33 males, 27 females; mean age 51.55 ± 13.58 years, range 8–74 years) and 22 PCNSL patients (14 males, 8 females; mean age 55.18 ± 12.19 years, range 32–78 years) were included in this study. All the PCNSLs were of the diffuse large B-cell type, and all patients were immunocompetent.

Results

The variables Firstorder_Skewness, Firstorder_Kurtosis, and Ngtdm_Busyness, representing features extracted from contrast-enhanced T1-weighted images, showed high discriminatory power. Firstorder_ Skewness was the best selected predictor for classification (AUC = 0.86), followed by Ngtdm_Busyness (AUC = 0.83) and Firstorder_Kurtosis (AUC = 0.80). The sensitivities and specificities ranged from 70.0% to 83.3% and from 71.4% to 90.5%, respectively. Among three classification models, the naive Bayes classifier was superior overall, with a high AUC (0.90) and the best specificity (0.91). The support vector machine models provided the best sensitivity and accuracy (0.92 and 0.88, respectively).

Conclusions

MRI-based 3D texture analysis has potential utility for preoperative discrimination of GBM and PCNSL.

Introduction

Glioblastomas(GBMs) and primary central nervous system lymphomas (PCNSLs) are common intracranial malignant tumors that share many visual characteristics on magnetic resonance imaging (MRI) [1,2]. Preoperative differentiation of GBMs and PCNSLs is of great clinical importance because the therapy and prognosis for the two entities are markedly different. For instance, GBMs are almost always treated by surgical intervention followed by radiation therapy and chemotherapy, whereas the treatment of choice for PCNSLs is methotrexate-based chemotherapy after a stereotactic biopsy [1,2]. In addition, preoperative steroid use usually complicates definitive histopathological diagnosis [1]. As a result, preoperative differentiation of GBMs and PCNSLs is a significant problem in need of a solution.

Although preliminary studies suggest that advanced MRI may be valuable in the task of distinguishing brain tumors, conventional MRI remains the preferred method for preoperative assessment in clinical practice [[3], [4], [5], [6], [7], [8], [9]]. However, the value of qualitative information obtained from conventional MRI is often limited. For example, atypical solid enhanced glioblastomas without visible necrosis may appear very similar to typical PCNSLs, making it difficult and sometimes even impossible to distinguish the two by qualitative methods (Fig. 1). Thus, objective quantitative methods based on conventional MRI may be clinically useful as a supplement to qualitative methods.

Texture analysis is a well-established technique to quantify parameters of a set of pixels (2D) or voxels (3D), i.e., the gray-level intensity and the distribution and relationship of image texture. Texture and shape analysis have been widely applied to medical images in recent years [10,11]. A considerable amount of research interest has been directed towards texture and shape analyses, which have attracted particular attention in the field of neurology, where these methods have demonstrated promising results [[12], [13], [14], [15], [16]]. Furthermore, some recent studies have concluded that heterogeneity information obtained from the total tumor mass (achieved by multiple MR image slices) is more valuable than that obtained from a single slice in texture analysis [[17], [18], [19]]. Considering the high rate of imaging-based misdiagnosis between GBM and PCNSL, we hypothesized that MR-based 3D texture and shape analysis would be able to aid in differentiating between GBM and PCNSL. Accordingly, machine-learning models were built to improve discrimination between these conditions using the results of this analysis.

Section snippets

Patients

We performed a database search of radiology and pathology reports from our Hospital between January 2012 and September 2017. Patients with solitary brain tumors shown on conventional MRI were identified by evaluating the MR image reports and postoperative pathological diagnoses. The histopathological diagnoses were determined according to the 2007 World Health Organization (WHO) classification criteria for glioblastoma and PCNSL [22,23]. The exclusion criteria included (1) non-enhancement or

Results

The three features with high predictive power were as follows: (1) Firstorder_Skewness, (2) Firstorder_Kurtosis and (3) Ngtdm_Busyness. The shape features showed no differences between glioblastomas and PCNSLs. However, according to the Mann-Whitney U test, the three above-mentioned parameters demonstrated statistically significant differences between GBMs and PCNSLs (p < 0.001, p < 0.001, and p < 0.001, respectively). Each individual ROC analysis of Firstorder_Skewness, Firstorder_Kurtosis and

Discussion

In this study, two common brain malignancies, GBMs and PCNSLs were differentiated by using MRI-based 3D texture and shape analysis. Three selected texture features with high predictive power were fed into the classification models. The results indicate that the texture feature analysis performs well and that the combination of the three machine-learning models aids in differentiation between the two tumors.

The first question in this study involves the efficacy of 3D texture analysis. According

Conclusions

MRI-based 3D texture analysis has potential utility in preoperative discrimination of glioblastomas and PCNSLs. In our study, the features with the highest predictive power in CE-T1WI were Firstorder_Skewness, Ngtdm_Busynessin, and Firstorder_Kurtosis, which had high sensitivity, specificity, accuracy and AUC. Machine-learning models using these three features also showed promising results, demonstrating their usefulness in the differentiation of glioblastomas and PCNSLs.

Funding sources

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Conflicts of interest

None.

Acknowledgements

None.

References (45)

  • A. Depeursinge

    Three-dimensional solid texture analysis in biomedical imaging: review and opportunities

    Med. Image Anal.

    (2014)
  • T. Batchelor et al.

    Primary CNS lymphoma

    J. Clin. Oncol.

    (2006)
  • N.A. Butowski et al.

    Diagnosis and treatment of recurrent high-grade astrocytoma

    J. Clin. Oncol.

    (2006)
  • J.G. Smirniotopoulos et al.

    Patterns of contrast enhancement in the brain and meninges

    Radiographics

    (2007)
  • P. Kickingereder et al.

    Primary central nervous system lymphoma and atypical glioblastoma: multiparametric differentiation by using diffusion-, perfusion-, and susceptibility-weighted MR imaging

    Radiology

    (2014)
  • A.C. Guo et al.

    Lymphomas and high-grade astrocytomas: comparison of water diffusibility and histologic characteristics

    Radiology

    (2002)
  • B. Hakyemez et al.

    Evaluation of different cerebral mass lesions by perfusion-weighted MR imaging

    J. Magn. Reson. Imaging

    (2006)
  • M. Law et al.

    High-grade gliomas and solitary metastases: differentiation by using perfusion and proton spectroscopic MR imaging

    Radiology

    (2002)
  • A. Larroza et al.

    Texture analysis in magnetic resonance imaging: review and considerations for future applications

    Assessment of Cellular and Organ Function and Dysfunction using Direct and Derived MRI Methodologies

    (2016)
  • M.E. Mayerhoefer et al.

    Texture-based classification of focal liver lesions on MRI at 3.0 Tesla: a feasibility study in cysts and hemangiomas

    J. Magn. Reson. Imaging

    (2010)
  • S.A. Waugh et al.

    Magnetic resonance imaging texture analysis classification of primary breast cancer

    Eur. Radiol.

    (2016)
  • K. Sogawa et al.

    Neurogenic and myogenic diseases: quantitative texture analysis of muscle US data for differentiation

    Radiology

    (2017)
  • Cited by (0)

    1

    These authors contributed equally to the work.

    View full text