Elsevier

Physica Medica

Volume 50, June 2018, Pages 26-36
Physica Medica

Original paper
Evaluation of radiomic texture feature error due to MRI acquisition and reconstruction: A simulation study utilizing ground truth

https://doi.org/10.1016/j.ejmp.2018.05.017Get rights and content

Highlights

  • MRI simulation was utilized to measure texture dependence on MR acquisition details.

  • Texture was measured on a digital brain phantom and clinical brain images.

  • Texture depends on noise, reconstruction algorithm, and parallel imaging R-factor.

  • Some texture features can be very different from the ground truth values.

  • Texture features are more or less robust and some predict high vs low grade glioma.

Abstract

The purpose of this study was to examine the dependence of image texture features on MR acquisition parameters and reconstruction using a digital MR imaging phantom. MR signal was simulated in a parallel imaging radiofrequency coil setting as well as a single element volume coil setting, with varying levels of acquisition noise, three acceleration factors, and four image reconstruction algorithms. Twenty-six texture features were measured on the simulated images, ground truth images, and clinical brain images. Subtle algorithm-dependent errors were observed on reconstructed phantom images, even in the absence of added noise. Sources of image error include Gibbs ringing at image edge gradients (tissue interfaces) and well-known artifacts due to high acceleration; two of the iterative reconstruction algorithms studied were able to mitigate these image errors. The difference of the texture features from ground truth, and their variance over reconstruction algorithm and parallel imaging acceleration factor, were compared to the clinical “effect size”, i.e., the feature difference between high- and low-grade tumors on T1- and T2-weighted brain MR images of twenty glioma patients. The measured feature error (difference from ground truth) was small for some features, but substantial for others. The feature variance due to reconstruction algorithm and acceleration factor were generally smaller than the clinical effect size. Certain texture features may be preserved by MR imaging, but adequate precautions need to be taken regarding their validity and reliability. We present a general simulation framework for assessing the robustness and accuracy of radiomic textural features under various MR acquisition/reconstruction scenarios.

Introduction

Radiomics, which regards images as data rather than pictures [1], involves, in part, the exploitation of information that cannot necessarily be discerned on an image, or set of images, by even an expert observer. These data are designed to be mined from standard radiologic images; thus, a shared multi-institutional database can lead to a large subject pool from which image-based predictive models of patient outcome can be built. Texture features are an important subset of the quantitative image characteristics that can be extracted, in addition to first-order intensity histogram statistics, shape-based features, and higher order statistical methods such as fractal and wavelet analysis. As a radiomics tool, texture analysis (TA) is increasingly being applied to diverse imaging modalities to provide anatomical segmentation, cancerous lesion delineation, and prediction of response of normal and pathologic tissue to radiation therapy [2], [3], [4], [5], [6]. TA of magnetic resonance (MR) images has been applied to diverse clinical sites in the context of radiation treatment (RT), for example: brain [7], [8], head and neck [9], [10], breast [11], kidney [12], bladder [13], prostate [14], [15], [16], [17], and extremities [18]. The results of those studies, as well as those of ongoing clinical research, indicate great potential for the incorporation of information gleaned from quantitative texture features into initial RT planning as well as adaptive RT of cancer patients.

Successful translation of quantitative imaging research into the clinic will depend in part on the ability to reliably repeat (e.g., multiple scans on the same subject) and reproduce (e.g., on various MR scanners) extracted texture features from MR images [19]. As noted by Mayerhoefer et al. [20], an impediment to widespread clinical application of MR-based TA is its sensitivity to the choice of MRI scanner and imaging protocol; several studies [14], [20], [21], [22], [23], [24] have investigated the dependence of TA on MRI field strength, scanner manufacturer, and MRI acquisition parameters in both living subjects and phantoms. While evaluation of repeatability and reproducibility of texture features measured by MRI are very important, in this paper we endeavor to address the issue of accuracy as well. There have been, to our knowledge, no attempts at TA validation by direct comparison of measured TA features to known ground truth texture of the object of interest. In the absence of absolute knowledge of the texture features of an object, only relative comparisons can be made between MRI-derived features extracted using various imaging protocols or multiple MR scanners. Moreover, if two MR imaging methodologies produce different texture features, then in the absence of a ground truth, one cannot determine which method is more accurate.

Our overall goal is to overcome the deficiencies of the aforementioned relative TA comparisons by creation of a ground truth digital MRI phantom. By utilizing the digital phantom as input to an appropriate MR image simulator, we are able to create images inside a wide universe of MR imaging scenarios, and establish the dependence of the absolute accuracy of extracted TA features on MR field strength, MR pulse sequence, arrangement of receive coils, presence of image artifacts, and choice of image reconstruction algorithm. The purpose of this paper is: i) to lay the basic methodological framework for determination of absolute texture feature dependencies, using a two-dimensional digital ground truth phantom as input to an MR simulator; ii) to evaluate the sensitivity of texture feature accuracy and variance to noise level, acceleration factor and image reconstruction algorithms; and, iii) to compare the magnitude of texture feature error and variance due to acquisition details and/or reconstruction algorithm choice to the magnitude of clinically relevant texture features differences (high versus low grade glioma) in clinical brain MR images.

Section snippets

Study design

Texture features were examined in i) simulated MRI datasets, based on a digital phantom of the brain, and ii) clinical brain MR images of glioma patients. For the first part of the study, an idealized 2D digital phantom of an axial slice of human brain, developed by Guerquin-Kern et al. [25] served as a ground truth. We employed an MRI simulator [26], developed by the same group that created the idealized brain phantom, that produces complex k-space data from the digital phantom. This simulator

Simulated data

Shown in Fig. 2 is the SER of images reconstructed using all four reconstruction algorithms, with zero added noise and no parallel imaging acceleration, as a function of refinement factor. Refinement factor is the degree of upsampling of the object to create k-space signal in these rasterized simulations. For example, a refinement factor of 3x means that the signal in each pixel of the 256 × 256 k-space that is reconstructed to create a 256 × 256 MR image, was calculated from a 768 × 768

Discussion

MR imaging, a well-established diagnostic modality due to its rich variety of tissue contrast, is being increasingly utilized for radiation therapy planning, image-guidance, and adaptive re-planning. Use of MR images in radiomics is complicated by the fact that MR image intensities do not generally reflect physical parameters (unlike electron density in CT), but vary depending on voxel size, magnetic field strength, pulse sequence, machine vendor and reconstruction algorithm [36]. Thus, the

Conclusion

Quantitative metrics derived from texture analysis of MR images created from a ground truth phantom are dependent on choice of reconstruction algorithm as well as image signal-to-noise ratio and parallel imaging acceleration factor. Simulations indicate that, for some of the texture features, MR acquisition details may be a contributor to measured feature variance among patients. While the difference from ground truth of some texture features on simulated MR images are small, and smaller than

Acknowledgments

The authors would like to thank Rebecca Mahon of Virginia Commonwealth University for providing useful literature references on repeatability and reproducibility of texture features.

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

References (42)

  • Aerts HJWL, Velazquez ER, Leijenaar RTH, Parmar C, Grossmann P, Carvalho S, et al. Decoding tumour phenotype by...
  • P. Brynolfsson et al.

    ADC texture-an imaging biomarker for high-grade glioma?

    Med Phys

    (2014)
  • M. Dang et al.

    MRI texture analysis predicts p53 status in head and neck squamous cell carcinoma

    AJNR Am J Neuroradiol

    (2015)
  • J. Fruehwald-Pallamar et al.

    Texture-based analysis of 100 MR examinations of head and neck tumors - is it possible to discriminate between Benign and malignant masses in a multicenter trial?

    Fortschr Röntgenstr

    (2015)
  • S.A. Waugh et al.

    Magnetic resonance imaging texture analysis classification of primary breast cancer

    Eur Radiol

    (2016)
  • A.S. Kierans et al.

    Textural differences in apparent diffusion coefficient between low- and high-stage clear cell renal cell carcinoma

    AJR Am J Roentgenol

    (2014)
  • Peng Y, Jiang Y, Antic T, Giger ML, Eggener S, Oto A. A study of T-2-weighted MR image texture features and...
  • A. Vignati et al.

    Texture features on T2-weighted magnetic resonance imaging: new potential biomarkers for prostate cancer aggressiveness

    Phys Med Biol

    (2015)
  • A. Wibmer et al.

    Haralick texture analysis of prostate MRI: utility for differentiating non-cancerous prostate from prostate cancer and differentiating prostate cancers with different Gleason scores

    Eur Radiol

    (2015)
  • R. Stoyanova et al.

    Prostate cancer radiomics and the promise of radiogenomics

    Translational Cancer Res

    (2016)
  • M. Vallieres et al.

    A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in soft-tissue sarcomas of the extremities

    Phys Med Biol

    (2015)
  • Cited by (85)

    • 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