MRI scans of the lumbar intervertebral segments (L2-S1) and trunk/core muscles of thirty-six (36) subjects (20 males and 16 females) were included in this study. All subjects were asymptomatic and older than 19 years of age. All subjects were scanned on a 3 T scanner using a standardized T2 weighted protocol.

None of the subjects reported a history of activity-limiting chronic back or leg injuries, nor experienced any low back pain at the time of the MRI scan. In accordance with Lee et al.’s (1988) exclusion criteria,”Potential participants who had (1) Degenerative changes in the lumbar spine (e.g., crushed vertebral body, trauma, etc.) and/or Erector Spinal muscles (ESMs) (e.g., atrophy); (2) Obvious spinal deformities; or (3) Any known pathology relevant to and likely to alter low back geometry (e.g., scoliosis and tumor) were not included in this study. Subjects were consented in accordance with the Institutional Review Board (IRB) at Auburn University.

Lumbar MRI scans (L2-S1) of the subjects were performed using a 70 cm Open Bore 3 T scanner (MAGNETOM Verio, Siemens AG, Erlangen, Germany) at the Auburn University MRI Research Center. All subjects were examined in a head-first-supine position (head towards the magnet with hands lying across their abdomen) with arm supports (available upon request) and leg support (mandatory). MRI coil selection included the in-table spine coils and a flexible body coil placed in between the pubic bones at the middle-lower anterior region of the abdomen. The flexible body coil was secured with straps to maximize signal consistency. The imaging protocols included a standard morphological T2-weighted turbo-spin-echo (TSE) sequence in the sagittal plane, Axial Continuous T2-weighted TSE, and Axial Multi-group T2-weighted images with the following parameters:

All subject survey and MRI data were anonymized and linked using a unique subject ID unrelated to their personal information. Data were archived and stored on a secure server.

Two researchers with experience collecting low back MRI images collected and measured all of the MRI data. One of the operators performed a localizer scan to verify subject placement (aligned straight on the scanner). A localizer scan was obtained to assist in subject placement and allow the analyst to focus on relevant regions of interest (see Fig. 1).

The two MRI analysts reviewed the scans separately. Data were collected from two studies [15, 16] (n = 26 and n = 10, respectively). Images were analyzed a second time after one month. The order of presentation of images was randomized for each analysis. Both analysts were blinded to subject identity. During the MRI interpretation, which considered all MRI sequences, the operators could freely adjust image brightness, contrast, and zoom, to select the slices that they felt were best for measuring the parameters of interest (e.g., disc size, muscle lever arms, etc.).

In order to evaluate MRI reliability, repeated scans with short inter-scan time intervals were performed. Reliability for the entire process was evaluated using a worst-case scenario that compared two distinct scans of the same subject with different analysts positioning subjects and performing each MRI scan (imaging protocol is shown in Fig. 2). In addition, the analysts measured each scan using Osirix^© software; their own scans, and the scans performed by their colleague.

Axial and Sagittal MRI scans were analyzed using an open-source digital imaging and communications in medicine (DICOM) software, OsiriX, 8.0.1, 32bit, (Bernex, Switzerland). At each spinal level from L2 to S1, 14 measurements were performed and shown in Figs. 3, 4, and 5. Abbreviations are used for these 14 measurements and are shown in Table 1.

The agreement between the examiners (inter-rater) and within each examiner (intra-rater) was analyzed using intra-class correlation coefficients (two-way mixed models: ICC_3,1) where analysts were fixed variables and subjects were random variables. Muscle Cross-sectional Area (CSA) and Intervertebral Disc (IVD) measurements were stacked together (from L2 to S1 level) to estimate an overall ICC value rather than estimating an individual score for each IVD level. Combining these measurements from multiple levels of the spine (L2 to S1), we obtain a more comprehensive assessment of the overall relationship between muscle CSA and IVD characteristics. In addition to this, aggregating data across multiple levels increases the sample size and statistical power of the analysis. These analyses were performed in SPSS (version 19). Bland–Altman plots were drawn to better visualize the data.

Mean Absolute Percent Differences (MAPDs) were calculated as: MAPD=∑i=1nXi-YiXi+Yi/2n×100 where Xi is the first measurement, Yi is the second measurement, n is the total number of observations.

Since Xi and Yi are two distinct measurements and there is no ground truth, the scores were averaged when estimating error to minimize bias.

In total, 144 MRI scans were obtained from 36 subjects. Table 2 presents the demographic data (age, sex, height, and weight). The average age was 23.7 years for males (SD 3.1) and 25.4 years for females (SD 4.8). At each lumbar level (L2-S1), measurements were taken for each IVD and Vertebral body by manually identifying and tracing the actual shape of the structures.

Two analysts measured all parameters three times with at least one month between repeated measurements of the same scan to assess the reliability and the repeatability of measurements. Data from six sets of measurements were compared. For each lumbar region analyzed (e.g., L4-L5), there were 50 distinct slices from which researchers could choose to perform their measurements. In order to test the reliability of the overall process, specific image slices were not pre-selected for the analysts prior to measurements. Each observer chose the slice they thought most appropriate for measuring the parameters. A comparison of different scans by operators and reviewed/measured by different analysts had not been previously conducted using a substantive sample size. The agreement of analysts in choosing the same slice or within one slice is shown in Table 3. The results show that the same slice was selected 61% of the time, and selections were within one slice over 90% of the time. The agreement was highest when researchers reanalyzed their own scans and was lowest but still excellent when they measured each other’s scans. Analysts waited at least one month between analyses of the same scans, with scans presented in random order each time.

Table 3 shows that when analysts repeated an analysis of a particular scan, they selected the same slice 79% of the time. Different analysts reviewing the same scan selected identical slices 50% of the time. Overall, it should be noted that analysts were within one slice of each other’s scans 90% of the time and did not differ by more than two slices (6 mm).

The mean absolute errors for each IVD level were reported as percentages in Tables 4 and 5. The results show that bony structures have lower error percentages except for AIVDH and PIVDH. In general, the smaller structures had bigger percentage errors. Moreover, when the analysts evaluated their own scans, error percentages decreased.

The ICCs for intra-rater reliability show high to excellent (0.806–0.989) agreement for both analysts and are reported in Table 6. The first sets of measurements for both analysts were compared to evaluate inter-rater agreement. The ICC results indicated high to excellent agreement for most of the measurements except PIVHD, SVBL, Disc Size, and CH. It should be noted that CH is a two-step measurement requiring analyst judgment [16], therefore can be more subjective then the other part measurements.

The Bland–Altman method for repeated measurement was performed to evaluate inter-rater reliability of the analysts for all regions. A representative Bland–Altman plot (Posterior Intervertebral Disc Height [PIVDH]) is shown in Fig. 6. A summary of the Bland–Altman analyses are in Table 7. The main purpose of using these plots is to visualize the agreement between analysts. In these plots, the Y-axis represents the difference between researchers’ readings, and the X-axis represents the average differences between analyst readings. Bias represents the mean of the differences in all measurements. The lower and upper levels of agreement (LOA) represent the mean difference’s 95% confidence interval (± 1.96 SD).

The lowest mean differences were observed in the Anterior Vertebrae Height (AVH) and the Sagittal Vertebrae Body Height (SVBH) by −0.03 cm. The highest mean differences were observed in the muscle groups (Psoas R, Psoas L, ESR, ESL) and the Disc Size. The highest mean differences are −0.30, −0.43, 1.99, 1.59, and −0.33 (cm²), respectively (shown in Table 7). A total of 14 measurements were made.

Positive correlations were exhibited between researchers for all measurements (Table 7). Some of these relationships were stronger than others. As complexity of tracing increases, so does the variability. Analysts have closer agreement in vertebral body measurements where the bony landmarks are highly visible. On the other hand, IVD and muscle CSA tracing is much more difficult. These, less well-defined structures were more difficult to measure and were more impacted by subject movement artifacts during scanning.