Figure 1.
An illustration of the algorithm.
(a) The spherical VOIlarge confines the search area. (b) The spherical VOImedium is defined for each pixel in VOIlarge to calculate CVv ( = standard deviation / mean). (c) VOI30 (sphere of 30-mm in diameter) is placed where CVv is smallest. The tumor delineation threshold is defined by the mean and SD within VOI30.
Figure 2.
A representative case processed with the semi-automated algorithm.
A 150-mm spherical VOIlarge is manually placed to roughly enclose the right lobe of the liver (a, b). For each voxel within VOIlarge, a spherical VOImedium (e.g., 80 mm in diameter) is defined, and mean (c), SD (d), and coefficient of variation of voxel (CVv) (e) within VOImedium are calculated. A 30-mm VOI30 is placed where CVv is minimized (f, g). Image color scales are 0 to 4 SUV for (a–c, f, g), 0 to 1 SUV for (d), and 0 to 0.25 (unitless number) for (e).
Figure 3.
To evaluate inter-study same-subject reproducibility of the VOI30 location, a cuboid region (red rectangle) was manually created to precisely contain the whole liver.
Location of VOI30 (black circle) was expressed as in the liver coordinate system.
Table 1.
Mean SUV values within-VOI30 from different methods and physicians.
Table 2.
Mean ±3 SD values within-VOI30 from different methods and physicians.
Table 3.
Number of successful results by semi-automated method.
Figure 4.
A representative case of two studies from the same patient.
(a, b) first study and (c, d) second study. The VOI30's defined using different size of VOImedium (blue: 40 mm, yellow: 60 mm, black: 80 mm, green: 100 mm, red: 120 mm) are drawn on transaxial slices (a, c) and maximum intensity projection images (b, d). The smaller VOImedium located VOI30 further from hepatic portal region with a larger distance between the first and second studies of the same patient than larger VOImedium's did.
Figure 5.
Euclidian distance between VOI30's from two subsequent studies of the same patient by different VOImedium sizes.
Except the combination of 100-mm and 120-mm, all the combinations showed significant difference (P<0.001) after Holm's correction for multiple comparisons.
Figure 6.
Bland-Altman plots of metabolic tumor volume.
A1 and A2 represent value from the semi-automated method operated by physician-1 and -2, respectively. M1 and M2 represent manually derived value by physician-1 and -2, respectively. (a) M1 vs. M2, (b) A1 vs. A2, (c) A1 vs. M1, and (d) A2 vs. M2 are compared. Solid lines represent mean difference and dashed lines represent mean ±2SD.
Figure 7.
Bland-Altman plots of relative change of the metabolic tumor volume (MTV), calculated as [MTVsecond – MTVfirst] / MTVfirst×100 (%).
A1 and A2 represent value from the semi-automated method operated by physician-1 and -2, respectively. M1 and M2 represent manually derived value by physician-1 and -2, respectively. (a) M1 vs. M2, (b) A1 vs. A2, (c) A1 vs. M1, and (d) A2 vs. M2 are compared. Solid lines represent mean difference and dashed lines represent mean ±2SD.
Figure 8.
Example images of CVv by different VOImedium's (a, d: 40 mm; b, e: 80 mm; c, f: 120 mm).
When 40-mm VOImedium was used (a, d), CVv seemed to be a non-convex function with many local minimums found in peripheral areas of the liver. When 80-mm VOImedium was used (b, e), CVv seemed to be a convex function. The minimum voxel existed in the right lobe of the liver. When 120-mm VOImedium was used (c, f), CVv seemed to be a convex function, but the minimum existed out of the liver. Image color scales are 0 to 0.25 for (a, d), 0 to 0.50 for (b, e), and 0 to 1.00 for (c, f).