Combined Volumetric and Density Analyses of Contrast-Enhanced CT Imaging to Assess Drug Therapy Response in Gastroenteropancreatic Neuroendocrine Diffuse Liver Metastasis

Objective We propose a computer-aided method to assess response to drug treatment, using CT imaging-based volumetric and density measures in patients with gastroenteropancreatic neuroendocrine tumors (GEP-NETs) and diffuse liver metastases. Methods Twenty-five patients with GEP-NETs with diffuse liver metastases were enrolled. Pre- and posttreatment CT examinations were retrospectively analyzed. Total tumor volume (volume) and mean volumetric tumor density (density) were calculated based on tumor segmentation on CT images. The maximum axial diameter (tumor size) for each target tumor was measured on pre- and posttreatment CT images according to Response Evaluation Criteria In Solid Tumors (RECIST). Progression-free survival (PFS) for each patient was measured and recorded. Results Correlation analysis showed inverse correlation between change of volume and density (Δ(V + D)), change of volume (ΔV), and change of tumor size (ΔS) with PFS (r = −0.653, P=0.001; r = −0.617, P=0.003; r = −0.548, P=0.01, respectively). There was no linear correlation between ΔD and PFS (r = −0.226, P=0.325). Conclusion The changes of volume and density derived from CT images of all lesions showed a good correlation with PFS and may help assess treatment response.

Liver is the most common site for GEP-NETs metastases; between 65 and 95% of GEP-NETs metastasize to the liver [6][7][8]. Diffuse liver metastases occur in most patients with GEP-NETs and directly influence prognosis in patients. Many treatment strategies can be applied to GEP-NETs with diffuse liver metastasis, of which medications play an important role in the management of unresectable liver metastases. Current standard medical treatment options include the use of somatostatin analogues, cytotoxic chemotherapy agents, and targeted agents [8].
Tumor treatment response can provide prognostic information and assist in determining the follow-up treatment strategy. Several metrics have been used in monitoring tumor response, among which chromogranin A (CgA) is currently recognized as the most valuable serum tumor marker used for screening, diagnosis, treatment, monitoring of progress, and prognostic evaluation [9][10][11][12]. However, there is no universal detection method, and the normal range and sensitivity of CgA vary according to the assay used [9][10][11][12]. is limitation restricts its application in current clinical practice. e use of computed tomography (CT) images is one of the most versatile and intuitive method [13]. RECIST [14] is widely used for the evaluation of treatment outcomes of anticancer therapy. However, RECIST criteria rely on diameter measurements on a single cross-sectional plane and on a limited number of lesions, making its interpretation controversial [15][16][17][18]. While some studies support its use for evaluating tumor response to treatment [15], others have suggested that it may overestimate or underestimate tumor burden [16][17][18]. Recent reports have claimed superiority of volumetric measurement of tumors over the use of RECIST [18][19][20]. However, a majority of these studies involved several representative lesions and did not reflect the change in all detectable hepatic lesions in patients with diffuse liver metastasis [18][19][20].
is study proposed a new computer-aided method to assess the treatment response using changes in total tumor volume and mean volumetric tumor density derived from volumetric measurements of all lesions in patients with GEP-NET diffuse metastases treated with drug therapy. e objective of this study was to assess the correlation of the changes in volumetric metrics with PFS that reflected the therapeutic effect. An analysis of its advantages and disadvantages over RECIST criteria is presented.

Patient Population.
e study protocol was reviewed and approved by the Ethics Committee of Sun Yat-Sen University, and informed consent was obtained from all patients. e methods were carried out in accordance with the approved guidelines. A total of 45 patients with GEP-NET diffuse liver metastases, who received at least two courses of drug therapy after resection of the primary tumor between February 2012 and June 2015, were retrospectively collected for this study. Subsequently, further screening was performed according to the inclusion criteria. Inclusion criteria were: availability of preand posttreatment dual-phase contrast-enhanced spiral CT images and progression-free survival (PFS), and absence of extrahepatic metastases. Twenty patients were excluded from the study due to either lack of data on posttreatment PFS (N � 4), lack of either pre-or posttreatment CT images (N � 10), or due to other extrahepatic disease that affected treatment outcomes (N � 5); one patient had unqualified CT images. Finally, a total of 25 patients (18 males and 7 females) came within the purview of this retrospective analysis, of which 10 cases had primary tumors located in the pancreas, 3 cases were located in the stomach, 7 cases were located in the small bowel, and 5 cases were located in the rectum (Table 1).

CT Imaging Protocol.
All patients underwent pre-and posttreatment CT imaging according to the standard institutional protocol for imaging of gastroenteropancreatic neuroendocrine liver metastasis. e posttreatment CT was performed after the end of two courses of treatment. A 64row spiral CT (Toshiba Aquilion64, Japan) equipment was used. Breath-hold unenhanced and contrast-enhanced images (matrix, 512 × 512; slice thickness, 1 mm; interslice gap, 0.8 mm) were obtained in the arterial (37 s) and portal venous phases (65 s). For contrast-enhanced CT, a dose of 1.5 ml/kg iopromide (Ultravist300, Schering, Berlin, Germany) was administrated at a rate of 3-4 ml/s. CT scan was obtained before contrast agent injection, 34-37 s and 60-70 s after contrast agent injection, respectively.

RECIST Measurement.
e follow-up observation was performed according to RECIST 1.1, which is based on the evaluation of a maximum of two target lesions per organ [21]. e sum of the longest diameters of the target lesions in each patient was computed in portal phase. en, the percentage change in tumor size from pretreatment levels was computed for each patient. e following were the definitions of the response criteria. Complete response (CR): disappearance of all target lesions. Partial response (PR): at least a 30% decrease in the sum of diameters of target lesions, taking as reference the baseline sum diameters. Progressive disease (PD): at least a 20% increase and an absolute increase of at least 5 mm in the sum of diameters of target lesions. Stable disease (SD) was defined as neither sufficient shrinkage to qualify for PR nor sufficient increase to qualify for PD. And, all patients were divided into two groups: progressive tumor (PD) and nonprogression tumor (PR or SD).

Progression-Free Survival (PFS).
Progression-free survival (PFS) is defined as the time between treatment initiation and evidence of tumor progression or death from any cause, with censoring of patients who are lost to follow-up or no tumor progression.

Volumetric Measurements.
In this study, a neural network classifier [22][23][24] was used to segment tumors from 3D 2 Contrast Media & Molecular Imaging CT liver images in the portal phase. Specifically, our whole segmentation framework consisted of two components ( Figure 1): (1) liver segmentation and (2) tumor segmentation from liver region. We adopted [22] to conduct liver segmentation, and combined [23] and [24] to segment tumor from liver region. In the first step, the liver region was identified by threshold method based on intensity analysis and anatomical knowledge [22]. We formulated the intensity distribution of CT images as a Gaussian mixture model with two components that represented liver region and nonliver regions, respectively. A large middle slice of liver was firstly segmented by a CT radiologist manually. en, the intensity range of the liver region was estimated by analyzing the statistical parameters of the Gaussian mixture model using expectation-maximization algorithm in this slice. e estimated intensity range was then used to threshold the images for liver segmentation. In order to discard components with a similar intensity as that of liver, we further refined the liver segmentation by keeping the largest connective region while discarding other regions based on the anatomical knowledge. And a morphological closing operation was performed to remove small holes in the segmented liver region. en, in the second step, within the segmented liver region, a backpropagation neural network-based classification method was employed to segment the tumors based on a series of grayscale co-occurrence features and statistical features [23,24]. e backpropagation neural network consisted of four layers, that is, one input layer, two hidden layers, and one output layer, which was trained using eight sets of CT liver images. For each voxel in the liver region, we extracted the statistical features and co-occurrence matrixbased features (e.g., average intensity, entropy, contrast, correlation [23]) within its 11 × 11 × 11 neighborhood. ese features were then fed to the backpropagation neural network for classification of the voxel as tumor or nontumor.
en, morphological closing and opening operations were performed to remove small holes and small noises, respectively. Finally, the tumor segmentation results were further refined by a CT radiologist. e following metrics were calculated based on the segmentation results: (1) total tumor volume: the sum of the tumor volumes from each segmented tumor was calculated for each patient, and the absolute and percent change in the sum from pretreatment levels computed; (2) mean volumetric tumor density: the mean CT value of all the tumor tissues as well as the absolute and percent change from pretreatment level was calculated for each patient.  Table 2. To investigate the value of total tumor volume and mean volumetric tumor density independently or in combination with treatment response, correlation analysis was performed. Correlation of Δ(V + D), ΔV, ΔD, and ΔS with PFS was assessed by Pearson correlation (r). Statistical analysis was performed using SPSS 20.0 software. P<0.05 was considered statistically significant.

Results
e number of hepatic lesions in an individual patient ranged from 6 to 129 before treatment and from 11 to 135 after treatment. e sum of the longest diameters of the target lesions of an individual patient according to RECIST criteria ranged from 1.4 to 19.5 cm (pretreatment) and from 2.0 to 23.5 cm (posttreatment). e total tumor volume of an individual patient ranged from 0.8 to 2168.0 cm 3 before treatment and from 1.7 to 2707.0 cm 3 after treatment. e mean volumetric tumor density in an individual patient ranged from 55.7 to 126.1 Hounsfield unit (HU) before treatment and from 50.7 to 143.5 HU after treatment.
According to the changes of the maximum axial diameter before and after treatment, 7 patients showed evidence of tumor progression, while the other 18 patients showed nonprogression (4 partial response and 14 with stable disease). In the tumor progression group, the tumor size increased by a mean of 175% after treatment. In contrast, the tumor size decreased by a mean of 13.5% after treatment in the tumor nonprogression group. For the tumor progression group, the mean increase in total tumor volume was 134.1%, but the mean decrease in volumetric tumor density was 8.7%. For the tumor nonprogression group, the total tumor volume increased by a mean of 36.6%, but the mean volumetric tumor density increased by a mean of 6.4%. Figures 2 and 3 show the tumor volume and HU histogram changes before and after treatment in one case of tumor regression and one case of tumor progression. Supplement video file 1 shows a 3D visualization of the segmented tumors in one patient.
Changes in total tumor volume and mean volumetric tumor density pre-and posttreatment for all 25 patients after treatment are shown in Tables 2 and 3. e total tumor volume and mean volumetric tumor density showed opposite changes in 6 patients (Table 3). Of 7 patients with tumor progression according to RECIST 1.1, 3 patients demonstrated increased total tumor volume and mean volumetric tumor density: the total tumor volume increased by 287.3%, 31.8%, and 81.3%, respectively, and the mean volumetric tumor density increased by 1.7%, 17.1%, and 11.0%, respectively. In another 3 patients with increased total tumor volumes and decreased mean volumetric tumor density, the total tumor volume increased by 296%, 131.2%, and 112.5%, respectively, while the mean volumetric tumor density decreased by 31.7%, 4.7%, and 24.2%, respectively. e remaining 1 patient demonstrated a decrease in both total tumor volume and mean volumetric tumor density by 1.6% and 30.2%, respectively. Of 18 patients with stable disease, 2 patients demonstrated decreased total tumor volume and mean volumetric tumor density: the total tumor volume decreased by 88.7% and 74.6%, and the mean volumetric tumor density decreased by 3.0% and 27.8%, respectively. 2 patients demonstrated increased total tumor volume and decreased mean volumetric tumor density, the total tumor volume increased by 57.9% and 43.5%, and the mean volumetric tumor density decreased by 14.5% and 13.1%, respectively. 2 patients with decreased total tumor volume and increased mean volumetric tumor density, the total tumor volume decreased by 28.7% and 27.5%, and the mean volumetric tumor density increased by 17.8% and 5.4%, respectively. 12 showed increased total tumor volume and mean volumetric tumor density.
During the study period, tumor progression was observed in 21 patients (84%). e remaining 4 patients (16%) exhibited no evidence of tumor progression. e median PFS of these 21 patients was 4.1 months (range 1.8-9.4 months).

Discussion
Our results demonstrated that the total tumor volume and mean tumor density obtained using volumetric measurements correlated better with PFS than tumor size of RECIST 1.1. Similar results have been reported elsewhere [18,25]. Hayes et al. reported superiority of volumetric measurements over use of unidimensional RECIST criteria to predict overall survival [25]. Welsh et al. also suggested volumetric analysis as the preferred method to detect tumor progression [18]. Just as some of the previous studies reported that RECIST might significantly underestimate or overestimate the tumor burden [16,18], our research suggested RECIST criteria were inadequate for a precise assessment of the tumor treatment response [17]. Being a unidimensional tumor metric, RECIST criteria were a valid surrogate for three-dimensional growth in tumors, only in the case of spherical-shaped tumors. However, liver metastases tend to have complicated shapes where it could not be accurately represented using RECIST metric.
Medications do not always cause tumoral shrinkage. Treatment response observed with targeted therapies could manifest as decrease in lesion size, decrease in lesion vascularity, cystic changes, and intratumoral hemorrhage, with or without a change in size [26]. is theory might also apply to nontargeted drugs. Reduction in tumor size is usually minimal during the early stages of treatment despite significant changes such as a decrease in the number of vessels [27]. erefore, morphological response assessment may not always provide an objective measure.
Functional imaging is increasingly being used for monitoring of response to anticancer therapy. CT enhancement correlates with tumor vascularity; so, the attenuation might indirectly reflect tumor activity. Several investigators had previously demonstrated that assessment of tumor density could improve therapeutic response assessment in metastatic gastrointestinal tumors [28,29]. However, at present, most of the studies have just measured the tumor CT density of different representative layers. Attenuation of selected level does not reflect the overall treatment response well.
ere is a growing recognition that intratumor heterogeneity could affect therapeutic effects in different areas of the same tumor [30].
Volumetric measurements of tumor response would reflect comprehensively the therapeutic effects of the various parts. Vargas et al. found a significant association between tumor grade and enhancement only when measuring the entire tumor and not on the most enhanced portion on a single slice [31]. Evaluation of whole-lesion attenuation had shown a better reproducibility than that of region of interest (ROI) measurements [32]. We measured the pre-and posttreatment mean volumetric density of the entire tumor and found no significant association between tumor density and PFS. However, combined tumor Contrast Media & Molecular Imaging 5 volume and density has been shown to better reflect therapeutic response. Smith et al. also found that evaluating changes in both tumor size and tumor density after targeted therapy remarkably improved the response assessment in metastatic renal cell carcinoma [33]. Our research demonstrated higher inverse linear correlation when combining change of total tumor volume and mean volumetric tumor density than using any one of these alone. is was also observed when compared with the change in tumor size measured using RECIST and PFS. For example, tumor necrosis might result in increase in volume and a concomitant decrease in density. erapeutic effect depends on the change in the amounts and activity of tumor tissue.
erefore, an analysis of the degree of change in total tumor volume and mean volumetric tumor density is more meaningful for purposes of treatment evaluation. e actual treatment response could be comprehensively determined using the percentage tumors, precise assessment of therapy response becomes more and more challenged. Our method might be improved to make it applicable to various usages in clinical imaging assessment for multiple metastases of different solid tumors, for example, lung cancer, breast tumor, and lymphoma. Furthermore, the popular deep learning methods could also be employed for more accurate and efficient organ and tumor segmentation in our future work.
ere were several limitations of our study. Firstly, due to the low overall incidence of GEP-NETs, there was a mixture of NET types with the primary in the pancreas, stomach, small intestine, and rectum, which may make it difficult to draw firm conclusions in heterogeneous patient group. Secondly, the sample size of our study is relatively small; to further study the superiority of volume and density measurement of tumors, a larger sample size is needed.    Contrast Media & Molecular Imaging irdly, the several different treatments applied also add to this problem of drawing firm conclusions from the present results.

Conclusions
We combined total volumetric and density analyses of contrast-enhanced CT imaging for assessment of therapeutic response in patients with GEP-NETs with diffuse liver metastasis. e volumetric and density analyses were carried out using semiautomatic segmentation of all tridimensional metastases. e combined use of total tumor volume and mean volumetric tumor density derived from the volumetric measurement may be helpful in the assessment of treatment response in patients with GEP-NETs with diffuse liver metastasis.

Abbreviations
GEP-NETs: Gastroenteropancreatic neuroendocrine tumors RECIST: Response Evaluation Criteria In Solid Tumors CgA: Chromogranin A ROIs: Regions of interest CR: Complete response PR: Partial response SD: Stable disease CT: Computed tomography HU: Hounsfield unit.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.

Authors' Contributions
Yi Wang, Kun Huang, and Jie Chen contributed equally to this work and should be considered as co-first authors.