Effect of obesity on ability to lower exposure for detection of low‐attenuation liver lesions

Abstract Purpose The purpose of this study was to assess the effect of obesity and iterative reconstruction on the ability to reduce exposure by studying the accuracy for detection of low‐contrast low‐attenuation (LCLA) liver lesions on computed tomography (CT) using a phantom model. Methods A phantom with four unique LCLA liver lesions (5‐ to 15‐mm spheres, –24 to –6 HU relative to 90‐HU background) was scanned without (“thin” phantom) and with (“obese” phantom) a 5‐cm thick fat‐attenuation ring at 150 mAs (thin phantom) and 450 mAs (obese phantom) standard exposures and at 33% and 67% exposure reductions. Images were reconstructed using standard filtered back projection (FBP) and with iterative reconstruction (Adaptive Model‐Based Iterative Reconstruction strength 3, ADMIRE). A noninferiority analysis of lesion detection was performed. Results Mean area under the curve (AUC) values for lesion detection were significantly higher for the thin phantom than for the obese phantom regardless of exposure level (P < 0.05) for both FBP and ADMIRE. At 33% exposure reduction, AUC was noninferior for both FBP and ADMIRE strength 3 (P < 0.0001). At 67% exposure reduction, AUC remained noninferior for the thin phantom (P < 0.0035), but was no longer noninferior for the obese phantom (P ≥ 0.7353). There were no statistically significant differences in AUC between FBP and ADMIRE at any exposure level for either phantom. Conclusions Accuracy for lesion detection was not only significantly lower in the obese phantom at all relative exposures, but detection accuracy decreased sooner while reducing the exposure in the obese phantom. There was no significant difference in lesion detection between FBP and ADMIRE at equivalent exposure levels for either phantom.

bias for lesion location by rotating the insert for different scans and to allow the creation of same-location lesion present/lesion absent pairs for a concurrent Channelized Hotelling Model Observer (nonhuman observer) study. The inserts were designed in quadrants, with lesions placed in different locations in each quadrant. In each insert, one to three quadrants contained one lesion (but not more than one) and one quadrant was always left blank.
Four LCLA lesions were studied based on the results of a previous published study 10

2.B | Scan technique and reconstruction
All scans were performed in helical mode on a Siemens Somatom Force Scanner (Siemens Healthineers, Forchheim, Germany) at 120 kVp without the use of automated exposure control because the phantom axial image attenuation profile is effectively uniform along the z direction; this also eliminated automated exposure control as a confounding factor.
The "thin" phantom was the standard anthropomorphic abdomen phantom (QRM) [ Fig. 2(a)]; this oval phantom is 30 cm wide by 20 cm in the anteroposterior dimension and 10 cm in length. Scans of this phantom were performed at 150 mAs (effective mAs, considered standard exposure for the thin phantom for this study), 100 mAs (33% exposure reduction), and 50 mAs (67% exposure reduction). The CTDI vol values for the three exposure levels were 9.9, 6.6, and 3.3 mGy × cm 2 , respectively.
The "obese" phantom was the standard phantom with the addition of a 5-cm thick ring of fat-attenuation material on the outside [ Fig. 2(b)], resulting in a phantom that was 40 cm wide by 30 cm in the anteroposterior dimension. Scans were performed at 450 mAs (effective mAs, considered standard exposure for the obese phantom for this study), 300 mAs (33% exposure reduction), and 150 mAs (67% exposure reduction). The CTDI vol values for the three exposure levels were 30, 20, and 10 mGy × cm 2 , respectively. These exposures, equal to 3 times the exposures used for the thin phantom, were considered clinically acceptable; to achieve a calculated noise in the center of the obese phantom that was equivalent to that achieved in the thin phantom would have required an approximately 7.4-fold increase in the exposure to >1100 mAs for full exposure, an exposure level that is not routinely used in or acceptable for clinical practice.
Each scan was performed a minimum of six times at each dose while the lesion insert was rotated among four 90-degree positions, moving the lesions within the phantom to minimize reader memory bias. The number of scans was determined by our statisticians to allow statistical significance at 10% accuracy loss for a noninferiority study using six readers. 10 Each scan was reconstructed with 3-mm slice thickness at 3-mm intervals, similar to abdominal CT scans performed at our institution, using both a B31f kernel for filtered back projection (FBP) and without and with iterative reconstruction (Adaptive Model-based Iterative Reconstruction (strength 3, ADMIRE, Siemens Healthineers). Scans were also reconstructed in both the craniocaudal and caudal-cranial directions to provide additional image datasets with variations of the noise pattern in the reconstructed images.   (Table 1).

2.E | Statistical analysis
A noninferiority model using Holm's step-down procedure was created to assess the accuracy of lesion detection with reduced dose for all lesions combined in both the thin and obese phantoms with FBP and ADMIRE. An accuracy loss of <10% was considered noninferior for this study. The Obuchowski-Rockette method for multireader multicase studies was used to compare the readers' mean receiver operating characteristic (ROC) areas.  T A B L E 1 Prevalence of lesions in the reader data set a by size and density (90-HU background).

3.B | Reader accuracy
With FBP, the mean area under the curve (AUC) was consistently lower for the obese phantom at each exposure level, even with the threefold increased exposure; at "full" exposure, for example, the AUC was 0.990 at 150 mAs for the thin phantom vs 0.916 at 450 mAs for the obese phantom. Similar results were seen with ADMIRE (e.g., mean AUC was 0.986 at 150 mAs for the thin phantom vs 0.914 at 450 mAs for the obese phantom) ( Table 4, Fig. 3).

3.C | Reader confidence
All six readers used the "indeterminate" score (3 on the 5-point Likert scale) at a higher frequency for the obese phantom than for the thin phantom at full exposure for both FBP (obese phantom, 13%; thin phantom, 8%) and ADMIRE (obese phantom, 14%; thin phantom, 8%).

| DISCUSSION
In this study, the accuracy for detection of LCLA liver lesions simu-  * P values adjusted for multiple comparisons using Holm's step-down procedure.

**
Comparisons that were not proven statistically noninferior (effectively inferior) (P > 0.05) for a 10% reduction in ROC area.
Iterative reconstruction preserved accuracy more at the lower exposures than higher exposures, as expected. While the improvement was not statistically significant, the trend suggests that when reducing exposure with CT, it is prudent to choose iterative reconstruction techniques over FBP for this type of lesion detection task.
Increasing radiation exposure for CT scans is necessary to maintain diagnostic efficacy in obese patients given the physics of CT. This was shown by Zhang et al 12  Iterative reconstruction techniques have been introduced as a method for reducing radiation exposure while limiting concomitant increases in noise, but these techniques have not been as effective at maintaining diagnostic efficacy as anticipated. In general, iterative reconstruction has at best modestly improved confidence in LCLA liver lesion detection without significantly improving the accuracy of detection. 10 Although one study showed subjective improvements in image quality with a noise-reducing iterative reconstruction algorithm, 13 several studies have shown little improvement in objective, task-based assessment of image quality with noise-reducing algorithms. 11,[14][15][16][17][18] One study showed increased detection with iterative reconstruction; however, this finding was for lesions 6 mm and smaller with low-contrast differentials. 15  None of these previous studies directly evaluated how obesity affects the ability to lower exposure and maintain LCLA liver lesion detection, either without or with iterative reconstruction. In this study, the decrease in accuracy (loss of noninferiority) of lesion detection with reduced exposure became significant sooner in the obese phantom than it did in the thin phantom, even at three times the exposure levels. Importantly, this study illustrates that exposure reductions in obese patients are more likely to result in lower diagnostic efficacy than they do in thin patients.
There are several limitations to this study. First and foremost, this is a phantom model of a liver lesion; the noise patterns in phantoms and human subjects are likely less uniform 21 and therefore the results may be different in human subjects, although likely worse. We also did not increase the exposure for the obese T A B L E 4 Direct comparison of accuracy for lesion detection in thin and obese phantoms. However, a previous study demonstrated no significant advantage with scrolling image assessment. 23 We also evaluated only lesion detection, not lesion characterization, but any degradation in the ability to characterize lesions would likely be greater than the ability to simply detect lesions. Finally, we studied only one CT scanner from one vendor, and thus the findings are not necessarily generalizable to other CT scanners or protocols, as has been shown previously. 12

| CONCLUSION
The results from this study using an anthropomorphic abdominal phantom show that subcutaneous fat plays an important role in the detection of LCLA liver lesions, requiring a markedly higher radiation exposure to achieve similar but still reduced detection rates at stan-