A comparative study of software programmes for cross-sectional skeletal muscle and adipose tissue measurements on abdominal computed tomography scans of rectal cancer patients

Abstract

Background: The association between body composition (e.g. sarcopenia or visceral obesity) and treatment outcomes, such as survival, using single-slice computed tomography (CT)-based measurements has recently been studied in various patient groups. These studies have been conducted with different software programmes, each with their specific characteristics, of which the inter-observer, intra-observer, and inter-software correlation are unknown. Therefore, a comparative study was performed.

Methods: Fifty abdominal CT scans were randomly selected from 50 different patients and independently assessed by two observers. Cross-sectional muscle area (CSMA, i.e. rectus abdominis, oblique and transverse abdominal muscles, paraspinal muscles, and the psoas muscle), visceral adipose tissue area (VAT), and subcutaneous adipose tissue area (SAT) were segmented by using standard Hounsfield unit ranges and computed for regions of interest. The inter-software, intra-observer, and inter-observer agreement for CSMA, VAT, and SAT measurements using FatSeg, OsiriX, ImageJ, and sliceOmatic were calculated using intra-class correlation coefficients (ICCs) and Bland-Altman analyses. Cohen's κ was calculated for the agreement of sarcopenia and visceral obesity assessment. The Jaccard similarity coefficient was used to compare the similarity and diversity of measurements.

Results: Bland-Altman analyses and ICC indicated that the CSMA, VAT, and SAT measurements between the different software programmes were highly comparable (ICC 0.979-1.000, P < 0.001). All programmes adequately distinguished between the presence or absence of sarcopenia (κ = 0.88-0.96 for one observer and all κ = 1.00 for all comparisons of the other observer) and visceral obesity (all κ = 1.00). Furthermore, excellent intra-observer (ICC 0.999-1.000, P < 0.001) and inter-observer (ICC 0.998-0.999, P < 0.001) agreement for all software programmes were found. Accordingly, excellent Jaccard similarity coefficients were found for all comparisons (mean ≥ 0.964).

Conclusions: FatSeg, OsiriX, ImageJ, and sliceOmatic showed an excellent agreement for CSMA, VAT, and SAT measurements on abdominal CT scans. Furthermore, excellent inter-observer and intra-observer agreement were achieved. Therefore, results of studies using these different software programmes can reliably be compared.

Keywords: Adipose tissue mass; Body composition; Comparability; Computed tomography; Skeletal muscle mass; Software.

Figure 1

Bland–Altman 95% limits of agreement plots for the agreement between the various software programmes (provided on the x‐axes and y‐axes) for CSMA (cm²). The dotted lines are the mean of the difference and the 95% limits of agreement (±2 SD) between the CSMA of reading 1 of observer A and the solid lines of reading 1 of observer B. (A) There was no proportional systematic bias for observer A (P = 0.908), whereas there was significant bias for observer B (P = 0.049). (B) There was no proportional systematic bias for any observer (P = 0.738 and P = 0.359). (C) There was no proportional systematic bias for any observer (P = 0.238 and P = 0.704). (D) There was no proportional systematic bias for any observer (P = 0.857 and P = 0.363). (E) There was no proportional systematic bias for any observer (P = 0.185 and P = 0.228). (F) There was no proportional systematic bias for any observer (P = 0.289 and P = 0.843).

Figure 2

Jaccard similarity coefficients (lowest and highest are shown) for inter‐software comparisons of CSMA, VAT, and SAT (cm²) measurements (reading 1 of observer B). The green area represents similarity, whereas the red area represents discrepancy in measurements. (A) The CSMA measured with FatSeg and ImageJ (1) and FatSeg and sliceOmatic (2), resulting in Jaccard similarity coefficients of 0.998 and 0.927, respectively. (B) The VAT measured with ImageJ and sliceOmatic (1, 2), resulting in Jaccard similarity coefficients of 0.998 and 0.855, respectively. (C) The SAT measured with ImageJ and sliceOmatic (1) and OsiriX and sliceOmatic (2), resulting in Jaccard similarity coefficients of 0.999 and 0.923, respectively.