Repeatability of Radiomic Features of Brown Adipose Tissue

Subjects

This study was approved by the Washington University institutional review board, and all participants provided written informed consent. Volunteers were recruited using flyers placed in various locations on the Washington University in St. Louis medical campus. Subjects received modest financial compensation for their investment of time in the study. Age and body mass index have been shown to be negatively correlated with BAT activity (5,20). Therefore, healthy adult volunteers aged 18 to 35 with a body mass index under 25 were eligible for this study. Pertinent exclusion criteria included active infections; current or past usage of nicotine products; a history of cold-related injury; insulin-dependent diabetes; current or past usage of medications that may interfere with BAT activity, such as β-blockers; and nursing or pregnancy.

The study consisted of 2 separate imaging sessions. During each session, participants underwent an individualized cooling procedure intended to activate BAT, followed by ¹⁸F-FDG PET/CT imaging. During the repeat session, participants underwent cooling and imaging using the same protocol. Each imaging session took place at the same time of day, and both sessions were completed within 2 wk.

BAT Activation Protocol

BAT activation was achieved using a previously reported protocol (19). Figure 1 shows a timeline of the imaging sessions. On arrival, participants changed into a hospital gown (clo value of 0.44) and were wrapped in warmed blankets for approximately 1 h before the start of the cooling procedure. The cooling procedure was initiated about 60 min before the planned time of ¹⁸F-FDG injection. Participants were outfitted in a cooling suit (CureWrap, MTRE; Advanced Technologies) that circulates chilled water. At the start of the BAT activation procedure, the cooling system was set at 10°C for all volunteers. The goal of the procedure was to cool participants as much as possible without inducing sustained shivering. When shivering was observed by the study team or reported by the participant, the temperature of the water was raised by 0.5°C every 60 s until shivering ceased. This cooling protocol continued from 60 min before ¹⁸F-FDG administration until approximately 10 min before ¹⁸F-FDG PET/CT imaging.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

BAT activation and imaging process. After a 1-h warming period, volunteers were cooled using a water-perfusion suit for approximately 2 h to activate BAT. After 1 h of cooling, ¹⁸F-FDG was administered and PET/CT imaging was performed immediately after the cooling procedure.

Imaging

PET/CT imaging was performed on a Siemens Biograph 40 (Siemens Healthineers). PET imaging was performed from the umbilicus to the base of the skull. About 60 min before the scan, each participant was approximately 185 MBq (5 mCi) (±20%) of ¹⁸F-FDG via intravenous bolus injection. A low-dose CT was performed immediately before PET imaging, using a tube potential of 120 kV, tube current of 80 mAs, pitch of 0.8, and rotation time of 0.5 s. PET imaging consisted of 3 contiguous, 8-min list-mode acquisitions. PET images were reconstructed using ordered-subset expectation maximization (OSEM; non-HD PET) and OSEM with point-spread function (HD PET).

Image Analysis

Because supraclavicular fossae generally contain the largest BAT depots in adult humans, this region was selected for radiomic analysis (21). The BAT segmentation process is illustrated in Figure 2. BAT segmentation was performed manually using MIM (version 6.7.7; MIM Software, Inc.) and according to a modified BARCIST 1.0 criteria: because all participants in the study were lean, a threshold based on SUV was used in place of SUV adjusted for lean body mass to segregate the activated BAT from the nonactive BAT and white adipose tissue. An SUV_max threshold of 1.2 or greater was applied to the supraclavicular fossae to identify potential areas of activated BAT. A CT Hounsfield unit (HU) range was then applied to help ensure the identified areas corresponded to fat.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2.

BAT segmentation process. Areas known to contain BAT (A) were manually delineated (B) and then thresholds were applied, using CT HU (C) and SUV (D) cutoffs, to obtain volumes of interest from which radiomic features were extracted (E and F).

PET image intensities were normalized to decay-corrected injected activity per kilogram of body weight (SUV [g/mL]), and CT images were left as HUs. Images were converted from Digital Imaging and Communications in Medicine (DICOM) to Nearly Raw Raster Data (NRRD), and the masks derived from supraclavicular BAT segmentation were converted from DICOM-radiotherapy structure set to NRRD format using 3D Slicer (version 4.8; https://www.slicer.org/) (22). Radiomic features were extracted using PyRadiomics (version 1.3; http://www.radiomics.io/pyradiomics.html), which complies with the imaging biomarker standardization initiative (IBSI) recommendations (23,24). Images were discretized to a fixed bin width of 0.1 SUV and 10 HU for PET and CT, respectively, as suggested by Leijenaar et al. (25). Given the sensitivity of textural features to anisotropy, images were resampled to 3-mm isotropic voxels using the Simple Insight Toolkit (ITK) B-Spline algorithm for interpolation (26). The number of analyzed radiomic features and wavelet decompositions per feature class are shown in Supplemental Table 1 (supplemental materials are available at http://jnm.snmjournals.org), and additional imaging details (per the IBSI reporting requirements) are provided in Supplemental Table 2. A total of 74 textural features that quantify spatial distribution of voxel intensities were extracted, including 23 gray level cooccurrence matrix, 14 gray level dependence matrix, 16 gray level run length matrix, 16 gray level size zone matrix, and 5 neighboring gray tone difference matrix features. Additionally, we extracted 19 first-order histogram and 14 shape features. A detailed description of each feature can be found on the pyRadiomics website (https://pyRadiomics.readthedocs.io/en/latest/features.html). The application of wavelet filters resulted in 8 decompositions per feature, considering all possible combinations of high-pass (H) or low-pass (L) filters in 3 dimensions (i.e., HHH, HHL, HLH, HLL, LHH, LHL, LLH, LLL). Only first-order and textural features were wavelet-decomposed (n = 744). These steps resulted in 851 unique radiomic features for each of the 3 image sets (i.e., HD PET, non-HD PET, and CT). As the volume of the segmented BAT in the right supraclavicular region was more consistent between visits compared with the left side (Lin’s concordance correlation coefficient [CCC] values of 0.669 and 0.474, respectively), radiomic features were extracted only from BAT depots on the right side.

Statistical Analysis

Lin’s CCC was used to assess radiomic feature repeatability (Eq. 1). The repeatability of radiomic features extracted from different imaging modalities and reconstruction methods was compared using the Wilcoxon signed-rank test. In all appropriate cases, Levene’s test was used to assess variance equality between groups, and variance heterogeneity between groups was remedied using reciprocal transformation. Radiomic feature family comparison and all subsequent analyses were performed using only features extracted from non-HD PET to avoid the additional impact of point-spread function filter incorporation. The repeatability of radiomic feature families was tested using the Kruskal–Wallis H test. The repeatability of radiomic features grouped by wavelet decomposition type was evaluated by the Friedman test and the Nemenyi post hoc test. Additionally, the effect of high versus low frequency decomposition on feature repeatability was separately assessed using the Wilcoxon signed-rank test. Eq. 1

where V1 and V2 indicate the first and second imaging visits, respectively, ρ is the Pearson correlation coefficient, σ is the radiomic feature SD from the indicated visit, and μ is the radiomic feature mean from the indicated visit.

Because SUV_max has been shown to be relatively consistent between test and retest imaging sessions, we hypothesized that radiomic feature repeatability would follow the trend in SUV_max relative difference (SUV_maxrd) between imaging sessions (Eq. 2) (19,27). To test this, subjects were stratified into a high-SUV_maxrd group and a low-SUV_maxrd group using a median SUV_maxrd cutoff. In order to assess the possible impact of changes in BAT metabolic volume (BMV) on radiomic feature repeatability, a similar analysis was performed by stratifying subjects into a high–BMV relative difference (BMV_rd) group and a low-BMV_rd group using a median BMV_rd cutoff. The CCC of each radiomic feature was then recalculated for both SUV_maxrd groups and both BMV_rd groups, and the Wilcoxon signed-rank test was used for groupwise comparison. Levene’s test was used to assess variance equality between groups. Eq. 2

where μ_V1,V2 specifies the mean of the visit SUV_maxV1 the visit SUV_maxV2.

The most informative BAT radiomic features were identified using cluster analysis. The most stable radiomic features extracted from non-HD PET (defined as those with CCC ≥ 0.8), as well as SUV_mean and SUV_max, were median centered, then cross-correlated. The coefficient of determination (R ²) between the features was computed, and features were then clustered together using hierarchical clustering with complete linkage and a Pearson correlation–based dissimilarity index (1-R ²). The cluster tree was cut at the height of 0.05 to designate clusters with highly correlated features (i.e., R ² > 0.95). This step allowed a dimensionality decrease and the identification of redundant features. The most informative radiomic feature within each cluster was found by calculating the normalized dynamic range (NDR) of each feature (Eq. 3). Equation 3

where n is the number of subjects in the group and R is the radiomic feature value from the indicated visit for subject i. R_max and R_min are the maximum and minimum radiomic feature values within the tested group, respectively.

All data analyses were performed in R (version 3.4.1; http://cran.r-project.org/). P values less than 0.01 were considered statistically significant, unless otherwise stated, and the Benjamini–Yekutieli method was used to control for multiple comparisons (28).