Original paperEvaluation of radiomic texture feature error due to MRI acquisition and reconstruction: A simulation study utilizing ground truth
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)
- et al.
Texture analysis of medical images
Clin Radiol
(2004) - et al.
Beyond imaging: the promise of radiomics
Phys Med
(2017) - et al.
Differentiation of true-progression from pseudoprogression in glioblastoma treated with radiation therapy and concomitant temozolomide by GLCM texture analysis of conventional MRI
Clin Imaging
(2015) - et al.
Characterization of texture features of bladder carcinoma and the bladder wall on MRI: initial experience
Acad Radiol
(2013) - et al.
MRI texture analysis on texture test objects, normal brain and intracranial tumors
Magn Reson Imaging
(2003) - et al.
Influence of multichannel combination, parallel imaging and other reconstruction techniques on MRI noise characteristics
Magn Reson Imaging
(2008) - et al.
Radiomics: the process and the challenges
Magn Reson Imaging
(2012) - et al.
Radiomics: images are more than pictures, they are data
Radiology
(2016) - et al.
The role of texture analysis in imaging as an outcome predictor and potential tool in radiotherapy treatment planning
Br J Radiol
(2014) - et al.
Predictive value of standardized intratumoral metabolic heterogeneity in locally advanced cervical cancer treated with chemoradiation
Int J Gynecol Cancer
(2016)
ADC texture-an imaging biomarker for high-grade glioma?
Med Phys
MRI texture analysis predicts p53 status in head and neck squamous cell carcinoma
AJNR Am J Neuroradiol
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
Magnetic resonance imaging texture analysis classification of primary breast cancer
Eur Radiol
Textural differences in apparent diffusion coefficient between low- and high-stage clear cell renal cell carcinoma
AJR Am J Roentgenol
Texture features on T2-weighted magnetic resonance imaging: new potential biomarkers for prostate cancer aggressiveness
Phys Med Biol
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
Prostate cancer radiomics and the promise of radiogenomics
Translational Cancer Res
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
Cited by (85)
Phantom study investigating the repeatability of radiomic features with alteration of image acquisition parameters in magnetic resonance imaging
2024, Journal of Medical Imaging and Radiation SciencesRadiomics: “unlocking the potential of medical images for precision radiation oncology”
2023, Machine Learning and Artificial Intelligence in Radiation Oncology: A Guide for CliniciansHarmonisation of scanner-dependent contrast variations in magnetic resonance imaging for radiation oncology, using style-blind auto-encoders
2022, Physics and Imaging in Radiation OncologyRadiomic feature reliability of amide proton transfer-weighted MR images acquired with compressed sensing at 3 T
2024, International Journal of Imaging Systems and Technology