Ethics and study population
Participant consent and approval from the Regional Committee for Medical and Health Research Ethics were previously obtained.
We obtained CT images of 41 subjects with T2DM enrolled in the Diabetes-study [29] and 91 healthy male subjects from the INFO-study [30]. The Diabetes-study population was aged 29 to 45 years (median 41), 49% males, with a mean body mass index (BMI) of 34.0 kg/m2. The INFO-study population was aged 38 to 45 years (median 40), all males, with a mean BMI of 26.4 kg/m2.
Computed tomography
Anonymised, unenhanced single abdominal CT images were acquired. In the Diabetes-study, CT images were obtained with a Somatom Volume Zoom, 4-slice CT scanner (Siemens Healthineers, Erlangen, Germany) at 5 cm above L4/L5 level in women and 10 cm above L4/L5 level in men with 120 kVp, 100 mAs and slice thickness 4 mm. In the INFO-study, CT images were obtained with a Somatom Sensation 64, 64-slice scanner (Siemens Healthineers, Erlangen, Germany) at L3/L4 level with 120 kVp, 200 mAs and slice thickness 5 mm.
Image analysis
Semiautomated body composition segmentation of the CT images was performed with the SliceOmatic software package (v 5.0 rev 7b, Tomovision, Magog, QC, Canada).
Body composition segmentation included four tissue compartments: abdominal muscle compartment (AMC), inter- and intramuscular adipose tissue (IMAT), VAT and SAT.
The segmentation was performed by four observers: three radiographers (radiology technicians) and one oncology resident. The resident had previously received training in the use of SliceOmatic at the University of Alberta hospital, Edmonton, AB, Canada. Over the course of 2 weeks, the resident held three 1-h teaching sessions for the technicians. This was followed by 7 to 12 h of practical training in the use of the software with individual feedback from the resident and two radiologists supervising the study.
Segmentation was performed according to the Alberta protocol, defined and used at the Alberta Hospital (AB, Canada), as shown in Fig. 1 [31]. By this definition, AMC is muscle tissue free of adipose tissue, not anatomical muscle which may include intramuscular fat, and IMAT was segmented separately as adipose tissue within the muscle fasciae. For each tissue, segmentation was restricted to the following predefined attenuation ranges: − 29 to 150 Hounsfield Units (HU) for AMC, − 190 to − 30 HU for IMAT and SAT and − 150 to − 50 HU for VAT [19, 31, 32].
The three radiographers (observers 1, 2 and 3) performed segmentation of all the CT images from both studies. They were organised into alternating pairs so that each image was analysed independently by two radiographers. To evaluate intraobserver variation, the three radiographers performed a second segmentation of the same images from the Diabetes-study, after a 1-month delay. The oncology resident (observer 4) performed segmentation of all images from the Diabetes-study. The observers were blinded to each other’s results and their own previous results. A flow chart describing the distribution of performed segmentations between the observers is shown in Fig. 2.
From the different modes, available in the SliceOmatic software, the observers utilised the Region growing-mode with the Paint and Grow 2D options. This mode allowed the users to delineate the different types of tissue based on predefined attenuation ranges. The tissue compartments were tagged in a specific order, starting with AMC followed by SAT, VAT and IMAT. Although the software could delineate these four compartments semiautomatically using the Grow 2D option, all the segmented compartments were manually adjusted in each image with the Paint tool to ensure that the compartments had been segmented correctly, especially around the muscle with nearby tissues of similar density such as bowels or kidneys, but also around vertebra and IMAT.
Exclusion criteria
CT images with inferior quality due to noise, respiratory artefacts or other movement artefacts were excluded from analysis. Images where AMC was cut from the field of view (FOV) bilaterally were also excluded. In images where the oblique or transverse abdominal muscles were cut unilaterally from the FOV, AMC was estimated by segmentation of the contralateral AMC multiplied by two. In images where SAT or VAT was cut unilaterally or bilaterally from the FOV, segmentation of the affected tissue compartment was not performed.
Statistical analysis
Segmentation data in square centimetres per tissue compartment for each image was exported from SliceOmatic to Microsoft Excel (version 14.0, Microsoft Corporation, Redmond, WA, USA) and analysed in IBM SPSS Statistics (version 23, IBM Corporation, Armonk, NY, USA) and R (version 3.3, www.r-project.org).
Descriptive statistics are presented as median, minimum, maximum and percentiles. Normal distribution of measurement data was evaluated with Q-Q plots and Shapiro-Wilk tests.
For each of the four tissue compartments, the underlying variations in the measurement data between individual subjects, individual observers and residual variation (random noise) were analysed with mixed effects models. Different levels of variation and interactions between variations were evaluated in three different potential mixed effects models named A, B and C. All models included variation between individual subjects and random noise. Additionally, model A included general observer to observer variation, model B included systematic variation between observers in how segmentation was performed and model C included both sources of variation.
Parameters for each model were estimated by restricted maximum likelihood. Model fit for the three models was evaluated by Akaike information criterion (AIC), and results from the best fitting model presented as estimated standard deviations with 95% confidence intervals.
Mixed effects models assume normally distributed residuals, which were not present in the measurement data as seen in the Q-Q plots. The efficiency of the applied estimation routine was evaluated by simulations, and a robustness analysis was performed on data transformed to achieve normally distributed residuals.
In addition to the mixed effects model estimates, intraclass correlation coefficients (ICCs) were calculated. Two-way random effect ICC was used for overall variation in segmentation results, and single measurement ICC for intraobserver variation. Confidence intervals for the two-way ICC were based on a random effects model with percentile bootstrap confidence intervals based on 10,000 replications randomly sampling subjects and observers.
Subgroup analyses were performed for the Diabetes-study and the INFO-study in order to explore differences in variations between the subject groups.