Elsevier

Magnetic Resonance Imaging

Volume 21, Issue 9, November 2003, Pages 989-993
Magnetic Resonance Imaging

Magnetic resonance imaging
MRI texture analysis on texture test objects, normal brain and intracranial tumors

https://doi.org/10.1016/S0730-725X(03)00212-1Get rights and content

Abstract

Texture analysis was performed in three different MRI units on T1 and T2-weighted MR images from 10 healthy volunteers and 63 patients with histologically confirmed intracranial tumors. The goal of this study was a multicenter evaluation of the usefulness of this quantitative approach for the characterization of healthy and pathologic human brain tissues (white matter, gray matter, cerebrospinal fluid, tumors and edema). Each selected brain region of interest was characterized with both its mean gray level values and several texture parameters. Multivariate statistical analyses were then applied in order to discriminate each brain tissue type represented by its own set of texture parameters. Texture analysis was previously performed on test objects to evaluate the method dependence on acquisition parameters and consequently the interest of a multicenter evaluation. Even obtained on different sites with their own acquisition routine protocol, MR brain images contain textural features that can reveal discriminant factors for tissue classification and image segmentation. It can also offer additional information in case of undetermined diagnosis or to develop a more accurate tumor grading.

Introduction

Several studies have addressed NMR tissue characterization for more than twenty years, both by ex-vivo relaxation times measurements and by quantitative MRI. Malignant intracranial tumors usually have longer relaxation times than benign tumors. However, the relaxation times overlap between these two kinds of tumors. Furthermore the different parts of the same tumor and often the surrounding edema are not clearly discriminated [1], [2], [3], [4], [5], [6], [7], [8], [9]. The average measurements performed in-vivo on full Region-of-Interest (ROI) or ex-vivo on surgery samples could explain these early disappointing results shading the tumor heterogeneity. Then a more sophisticated quantitative approach of MRI, taking into account the local pixel heterogeneity, was requested to improve the image segmentation and to contribute to a more clinically relevant tumor grading.

First used for satellite images, texture analysis (TA) medical applications first appeared in MRI at the beginning of the eighties [10]. TA covers several methods enabling quantification and statistical analysis of the distribution of the gray level values on an image ROI. A complete mathematical description of the TA parameters has already been published [11]. TA can offer information not visible by the human eye that can only discriminate complexity less than order two [12]. At this time, only very few studies have been performed to characterize brain tissue by TA. Lerski et al. [13], Schad et al. [14] and Kjaer et al. [15] have shown on preliminary clinical series that TA seems to be an interesting tool to characterize brain tissues and tumors such as glioblastoma and metastases in their general aspect (contrast, intensity, homogeneity) or their different constituents (micro or macro textures). They didn’t use raw data images (T1 and T2 weighted images) but T1 and T2 calculated images of metastases and heterogeneous brain tumors such as glioblastoma for their texture analysis.

In this study, a preliminary testing with TA test objects was first performed to evaluate the sensitivity and the effectiveness of the method in relation to variation of the MRI acquisition parameters. Comparatively to the previously published brain tumors TA studies performed on the same MRI unit with calculated T1 and T2 images, the goals of this study on a larger clinical sample size was to test the ability of TA to be performed: i) on more classic raw images (T1 or T2 weighted) and ii) on three different MRI units of three different university hospitals in their own routine acquisition conditions.

Section snippets

Test objects

Specific TA test objects were developed under the auspice of a European BIOMED Concerted Action (BMH-CT94-1274) [16]. They consisted of three tubes filled with gadolinium doped agarose gel (T1 = 360 ms and T2 = 80 ms at 1.5T and 296 K) and foams simulating coarse, medium or fine textures. The sizes of the foam porosities were respectively 1.69 to 3.70 mm for the first tube, 1.02 to 1.69 mm for the second one and 0.72 to 1.02 mm for the third one.

Patients

In agreement with the French ethical legislation

Texture analysis

Five circular ROI as large as possible (about 220 pixels) were chosen on three different slices and for each test object: 2 in the top gel and 3 in each foam to test the measurement reproducibility. On brain images, the ROI (were respectively 90 to 120 pixels and 40 to 180 pixels on the axial and sagittal slices according to the tissue studied) were visually defined as large as possible by a radiologist and chosen where the different parts (solid and liquid) of the tumor and the surrounding

Data analysis

Correspondence Factorial Analysis (CFA) was first performed with a specific software Bi@Loginserm [(1979–1987, INSERM, Paris (France)] using all the fifty texture parameters (2200 features for all subjects and 750 for all test tubes) to select the most statistically significant features (p < 0.05).

A Hierarchical Ascending Classification (HAC) was then used with the selected features to perform a four-class classification of test objects images: gel, large, intermediate and small porosities. For

Test objects

In the range of the SNR and resolutions tested for the gel and for the three foam sizes of the specific test objects used, no significant differences in TA classifications for both MPGR (T1w) and FSE (T2w) sequences have been evidenced.

The number of well-classified ROI (for the three foams sizes) was independent from the excitations number and slice thickness. The average number of well classified ROI was 6/10 for the gel, 8/10 for the fine foam, 8/10 for the medium foam and 9/10 for the coarse

Discussion and conclusion

TA method has been first presented as largely dependent from: i) the acquisition parameters, ii) the quality assessment of the MRI device, and iii) the methods of image reconstruction and processing [16], [17]. Then the application of texture analysis methods on three different MRI units using the same kind of MRI device but with slightly different sequence parameters, managed by different staffs for “maintenance” and for image acquisition was a challenge. The goal was to perform MR exams

Acknowledgements

We are grateful to J. Y. Bansard (Inst. Of public health, University of Rennes) for his helpful contribution in data analysis.

This work was supported by the Association pour la Recherche contre le Cancer (ARC) and by the Ligue Nationale Contre le Cancer.

References (18)

There are more references available in the full text version of this article.

Cited by (163)

  • Radiomics: “unlocking the potential of medical images for precision radiation oncology”

    2023, Machine Learning and Artificial Intelligence in Radiation Oncology: A Guide for Clinicians
View all citing articles on Scopus
View full text