A method for evaluation of patient-speci c lean body mass from limited-coverage CT images and its application in PERCIST: comparison with predictive equation

Jingjie Shang The First A liated Hospital of Jinan University Zhiqiang Tan The First A liated Hospital of Jinan University Yong Cheng The First A liated Hospital of Jinan University Yongjin Tang The First A liated Hospital of Jinan University Bin Guo The First A liated Hospital of Jinan University Jian Gong The First A liated Hospital of Jinan University Xueying Ling The First A liated Hospital of Jinan University Lu Wang The First A liated Hospital of Jinan University Hao Xu (  txh@jnu.edu.cn ) The First A liated Hospital of Jinan University https://orcid.org/0000-0002-6199-3929

measurement evaluated with LC images were much closer to the reference standard than those obtained by the James PE. The PE-based and LC-based PERCIST 1.0 classi cations were discordant in 27 patients (11.2 %; κ = 0.823, P=0.837). These discordant patients' percentage changes of peak SUL (SUL peak ) were all in the interval above or below 10 % from the threshold (± 30 %), accounting for 43.5 % (27/62) of total patients in this region. The degree of variability is related to changes in LBM before and after treatment.
Conclusions: LBM algorithm-dependent variability in PERCIST 1.0 classi cation is a notable issue. SUV normalised by LC-based LBM could change PERCIST 1.0 response classi cations based on LBM estimated by the James PE, especially for patients with a percentage variation of SUL peak close to the threshold. Background Quantitative 18 F-uorodeoxyglucose positron emission tomography ( 18 F-FDG PET) is expected to play a major role in assessing whether a tumour is responding to therapy, allowing physicians then to quickly determine whether to continue, change, or abandon treatment [1][2][3]. Fundamentally, quantitative PET treatment response assessment is based on the alteration in the standardized uptake value (SUV) between baseline and follow-up studies. But many technical factors affect SUVs and may lead to spurious variation [4]. To ensure the accuracy of assessment results, the PET evaluation response criteria in solid tumours (PERCIST) 1.0 have been proposed to mitigate these problems [5].
Nevertheless, there are still many sources of error in SUV measurement [6]. PERCIST 1.0 recommends using lean body mass (LBM) normalised SUV (SUL), due to its less variable between individuals of different body weights [5]. Modern PET/CT scanners commonly use the James predictive equation (PE), which relies on sex, height (cm) and total body weight (kg), to estimate LBM [7]. However, this equation has some limitations that weaken its reliability. A previous study has shown that the James PE was signi cantly different from LBM determined from dual-energy x-ray absorptiometry, which is one of the best reference methods [8]. In addition, Tahari et al. found inappropriately low hepatic SUL values in the very obese female patients when using this PE [9]. Therefore, a more accurate method should be proposed to ensure the reliability of LBM.
Computed tomography (CT) has become a standard method for measuring body composition [10].
Because different tissue types vary in x-ray attenuation on the basis of their different densities and chemical compositions, they can be distinguished on CT images. Thus, LBM can also be measured from the CT image obtained during the PET/CT examination [11]. However, conventional PET/CT examinations for most situations covered only the range from skull vertex to upper thighs and could not measure whole-body LBM. With this in mind, some have proposed using the limited-coverage (LC) CT to estimate LBM and demonstrated that LBM estimated from LC images has an excellent agreement with LBM measured on a whole-body CT [12,13].
Intuitively, more accurate measurement of LBM estimates from LC images should produce more accurate SUL normalisations than LBM estimates from the James PE. Narita et al. [14] reported that, compared with James PE-based SUL, individual CT-derived SUL could provide more stable hepatic uptake values that are available for PERCIST 1.0. However, the impact of the LBM algorithm on the PERCIST 1.0 classi cation of treatment response is unclear. PERCIST 1.0 suggests a threshold of 30% change in SUL (combined with a minimal absolute change) to de ne either partial response or progressive disease [5]. In practice, whether the difference between SUL normalisations would in uence the response classi cations, and how much the in uence would be, has not been reported. Therefore, the aims of this study are to introduce a novel and reliable method to estimate LBM by LC images from PET/CT examinations and compare the results with LBM estimates of the James PE. Then, analysis of whether SUV normalised by LC images-based LBM could change the PERCIST 1.0 response classi cations based on LBM estimated by the James PE.

Patients
First, a total of 296 patients who received whole-body 18 F-FDG PET/CT examinations between December 2011 and March 2018 were retrospectively reviewed. All patients were randomly assigned into equationdevelopment (106 men and 93 women) and validation (54 men and 43 women) groups. Then, patients with solid tumours who underwent 18 F-FDG PET/CT before and after treatment were retrospectively retrieved as the application group between the same periods. The inclusion criteria of this group were as follows: non-diabetic, still-existing 18 F-FDG-avid lesions and no appearance of new 18 F-FDG-avid lesions after treatment. Finally, a total of 241 patients (145 men and 96 women) were included in this study, including 72 with non-small cell lung cancer, 53 with hand and neck tumours, 46 with breast cancer, 22 with oesophageal cancer, and 48 with colorectal cancer. The clinic variables of patients were recorded, which included sex, age, height, weight, body-mass index (BMI), blood glucose level, injected dose of 18  Fat tissues were de ned as voxels identi ed and measured by CT as having CT numbers between −190 and −30 Houns eld units. A built-in software package of the Advantage Workstation (GE Healthcare) was used to calculate fat volumes (FV). A fat-tissue average density of 0.923g/mL was applied to convert whole-body FV to whole-body fat mass (FM WB ) [15]. LC images were obtained by truncating whole-body images. The LC region was de ned from top level of the thorax to distal point of the ischium. All image slices outside of this region were omitted. The volumes of LC fat tissues (FV LC ) were measured using the same method.
Traditionally, LBM was equivalent to the fat free mass (FFM) [13]. Because, in practice, the FM WB cannot be measured in conventional PET/CT examination, the relationships between FM WB and FV LC were analysed in the populations from the equation-development group, and an equation was developed using FM WB as the dependent variable and FV LC as independent variables. Subsequently, the LBM could be estimated from LC images (LBM LC ) according to the following equation: where α and β are the intercept and slope of the equation we developed, respectively.
To test the reliability of this method, LBM by LC images was compared with those derived by the James PE in an independent sample of 97 patients from the validation group. The reference standard was the measurement of LBM from whole-body CT. The James PE was de ned as follows: where W is weight in kg, and H is height in cm.

PERCIST evaluation
The therapeutic responses were analysed with PET volume computer-assisted reading (PET VCAR) of the Advantage Workstation (GE Healthcare). PET VCAR, using the James PE to estimate LBM and then measuring the peak SUL (SUL peak ) of target lesion (SUL peak-PE ), is one program the clinician can use to assist in monitoring treatment response according to PERCIST 1.0. This software can use an iterative adaptive segmentation algorithm to nd a threshold value that separated the target volume from the background tissue by weighting the maximum SUV (SUV max ) and mean SUV (SUV mean ) within the target volume with a default weighting factor of 0.5. A volume of interest (VOI) is placed over the tumor, and then the software automatically measures the SUL peak-PE within the entire tumor. Adjustment of the estimated tumor surface was sometimes needed to include the entire tumor within the margins of the volume of interest. The detailed instructions for PET VCAR were described in our previous study [16]. According to the LBM PE calculated by Equation 2 and the SUL peak-PE of target lesion measured by PET VCAR, the LC images-based SUL peak (SUL peak-LC ) of the same target lesion could be calculated.
As de ned in PERCIST 1.0 [5], the normal background region (a 3-cm diameter spherical region of interest [ROI]) was automatically delineated by PET VCAR in the right lobe of the liver. SUL peak of baseline lesion at least 1.5-fold greater than liver SUL mean + 2 SDs. If liver is abnormal, primary tumour should have uptake >2.0 × SUL mean of blood pool in 1-cm-diameter ROI in descending thoracic aorta extended over 2cm z-axis. Measurable lesion sizes should be 2 cm or larger in diameter for accurate measurement, though smaller lesions of su cient 18 F-FDG uptake, including those not well seen anatomically, can be assessed. We we chose the hottest lesion as the target lesion on the baseline and subsequent follow-up scans. The hottest lesion on the follow-up scan could be a lesion different from the previously measured lesion, on the assumption that it had been present since baseline. On the basis of the variation of SUL peak-PE and SUL peak-LC between the baseline and follow-up scans, patients were classi ed as partial metabolic response (PMR), stable metabolic disease (SMD), and progressive metabolic disease (PMD) separately according to PERCIST 1.0.

Statistical analysis
The patients' characteristics were presented as mean values and standard deviations. The unpaired t-test was used to analyse the differences in patients' characteristics between the equation-development and validation groups. For the equation-development group, Pearson correlation coe cients (r) were used to evaluate the relationship between the FV LC and FM WB . Then, simple linear regression was applied to generate equation to estimate FM WB from FV LC . The populations from the validation group were used for cross-validation: paired t-test and Bland-Altman plots were used separately to examine the difference and agreement between the outcomes of LBM PE and LBM LC and the reference standard.
The paired t-test was used to evaluate the differences in the same parameters between before and after treatment for the application group. Concordance and differences among the PERCIST 1.0 results of these two methods were assessed using Cohen's κ coe cient and Wilcoxon's signed-ranks test. Furthermore, to evaluate the impact of the LBM algorithm on assessment of the therapeutic response, we rst analysed the distribution of SUL peak-PE variation in patients with discordant response classi cations.
Subsequently, the variation of the LBM PE and LBM LC was evaluated in these patients. Graphs and analyses were performed using Prism GraphPad and the SPSS software.

Patient characteristics
Patients' clinical characteristics and parameters in the equation-development, validation and application groups are shown in   Table 2). The reason for the inconsistency in the evaluation results was due to the difference between the change rate of the PE-based SUL peak and the LC images-based exceeding the threshold.

Distribution of the percentage change of SUL peak in patients with discordant classi cations
The distributions of percentage change in SUL peak-PE for patients with discordant response classi cations were all within the interval above or below 10 % from the threshold (± 30 %). Over 10 %, although the variations of SUL peak-LC were different with SUL peak-PE , these differences were not enough to change the original classi cations. For all patients from the application group, a total of 62 patients' percentage change of SUL peak-PE distributed in this interval.
The percentage change of 62 patients' SUL peak-PE and SUL peak-LC are shown in Figure 2. Using 5 % interval as minimum unit, a total of 24 patients' percentage change of SUL peak-PE were found in the interval above or below 5% from the threshold, and 17 (70.8 %) discordances were found in this region. In addition, 38 patients' PE-based SUL peak were found in the interval above or below 5-10 % from the threshold, and 10 (26.3 %) discordances were found in this region.

Impact of LBM-dependent variation on PERCIST evaluation
Of 27 patients with discordant response classi cations, 16 patients' LBM showed a decrease at the follow-up examination, leading to PMR being changed to SMD and SMD to PMD. The reduction percentages of LBM PE (7.2 % ± 2.2) were signi cantly higher than those of LBM LC (2.8 % ± 1.8) (P<0.001).
Five patients' LBM showed increase, resulting in changes from SMD to PMR, and PMD to SMD. The increase percentages of LBM PE (5.7 % ± 1.5) were signi cantly higher than those of LBM LC (1.8 % ± 1.7) (P=0.001). Furthermore, six patients' weight increase and therefore LBM PE showed increase (1.8 % ± 1.0) at the follow-up examination, whereas LBM LC showed decrease (4.8 % ± 1.9), resulting in changes from SMD to PMR, and PMD to SMD.

Discussion
In this study, we proposed a novel and reliable method to estimate the patient's whole-body LBM from the LC images, and the results of this method showed an excellent agreement with the LBM obtained from the whole-body CT. Furthermore, and more importantly, we found that the classi cations of PERCIST 1.0 according to the SUL peak-PE could be reclassi ed according to the SUL peak-LC . The closer the percentage changes of SUL peak to the threshold, the greater the change that could have happened.
CT technology is considered one of the preferred approaches for measuring FM [10], but the application of this technology is limited by the radiation exposure. The previous studies demonstrated that the single abdominal image was highly correlated with the total volume of fat tissue (r=0.88-0.963) [17][18][19]. Accordingly, some investigators proposed, as a compromise between accuracy and reduced radiation burden, using a single cross-sectional image at the abdomen to predict whole-body FM [19,20]. In this study, we analysed the correlation between a wider range of LC fat volumes and whole-body FM. Similarly, the results showed that the FV LC were signi cantly correlated with the FM WB (r=0.977).
Compared with single-slice, the region of LC contained more fat tissues in the body and theoretically could enable estimation of FM WB more accurately.
Previous studies demonstrated that the LBM from the LC images was more accurate than the results obtained from the James PE. In a study by Chan [12], the LBM was estimated from the James PE and the LC images (at least from skull base to 5 cm below the pelvis). The results showed that the LBM was overestimated by the James PE, but not by the LC images method. The reliability of this method was con rmed in a later study [13]. Similar results were obtained in this study. We found that the LBM PE was signi cantly higher than the reference standard, but not for the LBM LC . The Bland-Altman plots showed that the results of LBM LC were more accurate than those of LBM PE .
A reasonable explanation for the heterogeneity of LBM PE is that the LBM has ethnicity speci city [21], the James PE was derived from Caucasian populations and was therefore not suitable to apply to a different ethnic group [22]. On the other hand, one major advantage of the method we proposed over the James PE is that it is based mainly on the directly measured LBM of an individual, albeit incomplete, rather than on an assumed similarity between the subject and some speci c study population. The accuracy of this method has been proved by the validation group. Furthermore, different from the method proposed by Chan [12], who estimated LBM by using the relative contribution of FM from a larger LC (from the eye to thigh), the method we proposed in this study concerned using the relationship between FV LC and FM WB to estimate LBM. The reason for selecting the region of the thorax to ischium is that the conventional PET/CT scans de nitely included this anatomic region, and the anatomical location is clear and easy to popularise.
Through the clinical application, we found that PE-based and LC images-based PERCIST 1.0 was discordant in 27 (11.2 %) patients. These discordances were due to differences in the values of SUL peak-PE and SUL peak-LC and, more important, to the variance in LBM between the baseline and follow-up scans. In general, patients who received cancer treatment suffered from weight loss, especially skeletal muscle loss [23,24]. In this study, 16 of 27 patients had a decrease in LBM after treatment, leading to the classi cations of PMR and SMD determined by PE-based PERCIST 1.0, and were reclassi ed as SMD and PMD by LC images-based PERCIST 1.0. These differences illustrate that the percentage of reduction of LBM PE was signi cantly higher than that of LBM LC (P<0.001), and the percentage change of SUL peak-PE was therefore smaller than that of SUL peak-LC . These could lead to worse classi cations when the percentage change of SUL peak-LC exceeded the thresholds. In addition, ve of 27 patients had an increase in LBM after treatment; conversely, the response evaluation based on LC images-based PERCIST 1.0 tended to be more optimistic, from SMD and PMD reclassi ed as PMR and SMD.
It is noteworthy that six of 27 patients had an increase in body weight and therefore in LBM PE after treatment, but LBM LC showed a decrease, resulting in more optimistic classi cations based on LC images-based PERCIST 1.0. These discrepancies are most likely explained by a state, for patients with chronic diseases, in which the loss of muscle mass was associated with preserved or even increased body fat content [25]. Multiple reasons for this state have been suggested, including received chemotherapy, decreased physical activity, increased total caloric intake, altered endocrine function and in ammation [25,26]. There could be marked weakness despite normal or even increased weight, causing the James PE derived from a population of normal health to produce a larger error.
Although heterogeneity was found only in the region close to the threshold, it is also crucial, because for individual patients, different response classi cations may lead to different treatment programs. To ensure the accuracy of response evaluation, PERCIST 1.0 describes in detail methods for controlling the quality of 18 F-FDG PET imaging conditions and provides a much more detailed framework for lesion selection, region of interest de nition, and response classi cation [5]. However, in this study, we found that on the basis of standardised PET/CT scanning process and quality control, accurate measurement of LBM is also essential to ensure the accuracy of evaluation. investigate the correlation of these discordances with pathological examinations to con rm this conclusion.

Conclusions
The whole-body LBM could be estimated from LC CT data acquired in PET/CT examinations. The patientspeci c LBM was more accurate than results obtained by the James PE, which has been routinely used in the modern PET/CT system. LBM algorithm-dependent variability in PERCIST  Figure 1 Bland-Altman plots of LBM computed using the James PE (a) and the method we developed (b) with LBMWB as reference standard.