Effect of grey-level discretization on texture feature on different weighted MRI images of diverse disease groups

Abstract

Purpose Many studies of MRI radiomics do not include the discretization method used for the analyses, which might indicate that the discretization methods used are considered irrelevant. Our goals were to compare three frequently used discretization methods (lesion relative resampling (LRR), lesion absolute resampling (LAR) and absolute resampling (AR)) applied to the same data set, along with two different lesion segmentation approaches. Methods We analyzed the effects of altering bin widths or bin numbers for the three different sampling methods using 40 texture indices (TIs). The impact was evaluated on brain MRI studies obtained for 71 patients divided into three different disease groups: multiple sclerosis (MS, N = 22), ischemic stroke (IS, N = 22), cancer patients (N = 27). Two different MRI acquisition protocols were considered for all patients, a T2- and a post-contrast 3D T1-weighted MRI sequence. Elliptical and manually drawn VOIs were employed for both imaging series. Three different types of gray-level discretization methods were used: LRR, LAR and AR. Hypothesis tests were done among all diseased and control areas to compare the TI values in these areas. We also did correlation analyses between TI values and lesion volumes. Results In general, no significant differences were reported in the results when employing the AR and LAR discretization methods. It was found that employing 38 TIs introduced variation in the results when the number of bin parameters was altered, suggesting that both the degree and direction of monotonicity between each TI value and binning parameters were characteristic for each TI. Furthermore, while TIs were changing with altering binning values, no changes correlated to neither disease nor the MRI sequence. We found that most indices correlated weakly with the volume, while the correlation coefficients were independent of both diseases analyzed and MR contrast. Several cooccurrence-matrix based texture parameters show a definite higher correlation when employing the LRR discretization method However, with the best correlations obtained for the manually drawn VOI. Hypothesis tests among all disease and control areas (co-lateral hemisphere) revealed that the AR or LAR discretization techniques provide more suitable texture features than LRR. In addition, the manually drawn segmentation gave fewer significantly different TIs than the ellipsoid segmentations. In addition, the amount of TIs with significant differences was increasing with increasing the number of bins, or decreasing bin widths. Conclusion Our findings indicate that the AR discretization method may offer the best texture analysis in MR image assessments. Employing too many bins or too large bin widths might reduce the selection of TIs that can be used for differential diagnosis. In general, more statistically different TIs were observed for elliptical segmentations when compared to the manually drawn VOIs. In the texture analysis of MR studies, studies and publications should report on all important parameters and methods related to data collection, corrections, normalization, discretization, and segmentation.

Phantom Study on the Robustness of MR Radiomics Features: Comparing the Applicability of 3D Printed and Biological Phantoms

“…In this study, we investigated the repeatability measures at different steps in the analyses, the normalization process, and the RI, in addition to how they were affected by object size and imaging parameters. We report improved repeatability for both biological and printed phantoms when normalization is applied to the data ( Supplementary Figures S1 and S2 , Table 4 ), findings that are in concordance with previous studies [ 7 , 52 , 56 , 94 , 95 , 96 , 97 ]. Similarly, when normalization was applied to the data, improvements were observed for the two discretization methods (FBS and FBN).…”

“…Several studies of biological phantoms have proven that the segmentation procedure may affect texture analysis; hence we chose two robust volume delineation methods: the fixed cube-shaped VOI definition and the “grow from seeds” 3D Slicer tool [ 7 , 47 ]. Figure 4 A demonstrates that the visible robust texture of the Hilbert cube is not affected by the spatial resolution when using 1 mm and 2 mm isotropic volumes.…”

“…The [] brackets stand for the ceil operation. We choose B = 0.15 for normalized images and B = 50 for non-normalized ones to have a similar number of bins in both cases [ 7 , 12 , 59 ].…”

“…RIs were extracted using the GLCM (Gray Level Co-Occurrence Matrix), GLSZM (Gray Level Size Zone Matrix), and GLRLM (Gray Level Run Length Matrix)-based algorithms implemented in MATLAB (version 2020) [ 7 ]. A total of 40 radiomic features were extracted from each VOI, divided into 18 GLCM, 11 GLSZM, and 11 GLRLM metrics.…”

“…First, the spatial resolution of the diagnostic imaging modalities (spatial resolutions of 0.5–5 mm) are orders of magnitude worse than the histological samples (spatial resolution of 10 −4 –10 −3 mm). Second, the choice of radiomic features may be affected by the diagnostic image modality and their modality-specific reconstruction settings [ 7 , 8 , 9 , 10 , 11 ]. Third, image discretization and normalization procedures may affect the radiomic analyses.…”