Volumetric and texture analysis of pretherapeutic 18F-FDG PET can predict overall survival in medullary thyroid cancer patients treated with Vandetanib

Purpose The metabolically most active lesion in 2-deoxy-2-(18F)fluoro-D-glucose (18F-FDG) PET/CT can predict progression-free survival (PFS) in patients with medullary thyroid carcinoma (MTC) starting treatment with the tyrosine kinase inhibitor (TKI) vandetanib. However, this metric failed in overall survival (OS) prediction. In the present proof of concept study, we aimed to explore the prognostic value of intratumoral textural features (TF) as well as volumetric parameters (total lesion glycolysis, TLG) derived by pre-therapeutic 18F-FDG PET. Methods Eighteen patients with progressive MTC underwent baseline 18F-FDG PET/CT prior to and 3 months after vandetanib initiation. By manual segmentation of the tumor burden at baseline and follow-up PET, intratumoral TF and TLG were computed. The ability of TLG, imaging-based TF, and clinical parameters (including age, tumor marker doubling times, prior therapies and RET (rearranged during transfection) mutational status) for prediction of both PFS and OS were evaluated. Results The TF Complexity and the volumetric parameter TLG obtained at baseline prior to TKI initiation successfully differentiated between low- and high-risk patients. Complexity allocated 10/18 patients to the high-risk group with an OS of 3.3 y (vs. low-risk group, OS = 5.3 y, 8/18, AUC = 0.78, P = 0.03). Baseline TLG designated 11/18 patients to the high-risk group (OS = 3.5 y vs. low-risk group, OS = 5 y, 7/18, AUC = 0.83, P = 0.005). The Hazard Ratio for cancer-related death was 6.1 for Complexity (TLG, 9.5). Among investigated clinical parameters, the age at initiation of TKI treatment reached significance for PFS prediction (P = 0.02, OS, n.s.). Conclusions The TF Complexity and the volumetric parameter TLG are both independent parameters for OS prediction.


Introduction
The tyrosine kinase inhibitor (TKI) vandetanib, a selective inhibitor of wild-type rearranged during transfection (RET) kinase as well as of vascular endothelial growth factor receptor (VEGFR) signaling [1][2][3], has demonstrated a favorable disease control in patients suffering from advanced medullary thyroid carcinoma (MTC) [4,5]. The increased use of vandetanib outside controlled clinical settings, as well as the attendant cost and potential toxicities of the drug, underscore the need for prediction of patients who are most likely to benefit from treatment [6][7][8].
In a recent study evaluating MTC patients prior to vandetanib, the prognostic potential of the metabolically most active lesion derived by baseline 18 F-FDG PET for prediction of progression-free survival (PFS) could be demonstrated. However, this simplistic strategy failed for overall survival (OS) [21]. Hence, in the present proof of concept study, we aimed to explore, if intratumoral heterogeneity textural features (TF) as well as volumetric parameters derived from 18 F-FDG PET prior to and 3 months after vandetanib initiation succeeded in prognostication in patients with advanced MTC.

Patient cohort
In this retrospective evaluation, a secondary analysis was performed of the identical patient cohort analyzed in [21].
Patient's characteristics are also provided in Supplementary  Table 1. All relevant clinical and outcome parameters were updated (including a novel date of censoring) and a complete re-analysis of the data was performed. All patients gave written informed consent to the diagnostic and therapeutic procedures as well as to the scientific analysis of the obtained data. Due to the retrospective nature of this analysis, the requirement for additional approval was waived by the local institutional review boards. Parts of this cohort received vandetanib in the context of a clinical trial [5].
A detailed description of the study cohort can be found in ref. [21]. In brief, between April 2007 and January 2018, 18 patients (6 females, median age at start of TKI initiation, 48 years) with advanced MTC receiving vandetanib (300 mg orally per day) were included. 14 patients were recruited at the University Hospital Würzburg, Germany, whereas the remaining 4 patients were treated at the Hospital of Augsburg, Germany. All patients had undergone various previous therapies (surgery in all patients; external beam radiation therapy in 4/18 (22.2%); chemotherapy in 3/18 (16.7%); transarterial chemoembolization in 2/18 (11.1%); sorafenib in 1/18 (5.6%) and radioiodine therapy in 1/18 (5.6%), due to an initial mis-classification as differentiated thyroid cancer). For further details refer to [21].

Radiological Response Assessment
Response assessment was performed on a 3-months basis according to Response Evaluation Criteria in Solid Tumors (RECIST) 1.1. The best response achieved by computed tomography (CT) was also investigated (i.e., Complete Response (CR), Partial Response (PR), Stable Disease (SD) and Progressive Disease (PD)) [22]. PFS covered the time span from vandetanib initiation to the time point of RECIST-based disease progression. OS (median) was defined using the following formula: [(Date of death)-(Date of treatment initiation)] [21].

Imaging
Imaging was performed on a stand-alone lutetium oxyorthosilicate full-ring PET scanner (ECAT Exact 47, Siemens Medical Solutions, Erlangen, Germany) in 4/18 (22.2%) of the patients (with a separate CT available in all), whereas 14/18 (77.8%) underwent integrated PET/CT (12/ 14, 85.7%, Biograph mCT PET/CT, Siemens Medical Solutions, Erlangen, Germany and 2/14, 14.3%, Gemini TF 16 PET/CT system, Philips Medical Systems, Hamburg, Germany). Prior to imaging, patients fasted for a minimum of 6 h (blood glucose levels < 160 mg/dl). 18 F-FDG was administered intravenously and 60 min post-injection, transmission data were acquired from the base of the skull to the proximal thighs using 68Ge rod sources (in the case of the stand-alone PET scanner) or spiral CT. Consecutively, the PET emission data were acquired. After decay and scatter correction, the PET data were reconstructed iteratively with attenuation correction, using the algorithm supplied by the scanner manufacturer. For further details, please refer to [21]. After 3 months, another 18 F-FDG PET/CT was performed in 16/18 (88.9%), CT in 1/18 (5.6%) and in one subject, imaging follow-up could not be obtained because of early termination of the treatment by the patient.

Image interpretation
In contrast to the previous study, which exclusively investigated a single region of interest drawn around the metabolically most active lesion [21], a complete re-analysis of the entire tumor burden of every patient was performed: the tumor volume of all tumor lesions was determined in a consensus analysis by two board-certified nuclear medicine physicians with long-standing PET/CT experience. Volumes of Interest (VOIs) were set by using a computerized 3-dimensional volumetric rendering tool (Interview Fusion Workstation, Mediso Medical Imaging Systems Ltd., Budapest, Hungary) [23]. CT images were not used to guide delineation of the VOIs [24]. Any non-tumoral, physiological areas of 18 F-FDG uptake were excluded. Within each VOI, the following conventional PET parameters were derived: maximum/mean standardized uptake value (SUV max/mean ), Metabolic Tumor Volume (MTV) and Total Lesion Glycolysis (TLG). MTV was defined by volume delineation and TLG was calculated using the following formula: [MTV x SUV mean ] [9,25]. All conventional PET parameters as well as TF were calculated automatically by the software. The radiotracer concentration in the VOIs was decay corrected and normalized to the injected dose per kilogram of patient's body weight to derive the SUV.
The following TF, which had been investigated in a previous study for outcome prediction in non-small cell lung cancer (NSCLC) patients under TKI treatment [19], were investigated: the first-order parameters Standard Deviation and Kurtosis, the second order parameters Entropy and Homogeneity as well as the higher-order parameters Busyness, Coarseness, Complexity, and Contrast. A detailed description can be found in [19]. In brief, first-order parameters inherit global textural features that relate to the gray level frequency distribution and are based on histogram analysis. Second-order parameters are derived from neighborhood spatial gray level dependence or cooccurrence matrices M. The matrix M determines how often a pixel with intensity i finds itself within a relationship to another pixel with intensity j in a VOI. Of note, the cooccurrence matrix describes only the changes of one voxel to the immediate next voxel. Higher-order parameters are calculated from 3-dimensional neighborhood gray-tone (intensity) difference matrices (NGTDM) to describe local features. These parameters are based from differences between each voxel and the neighboring voxels in adjacent image planes: The NGTDM (M4) contains entire homogeneous areas of a certain intensity and size [26]. Higherorder TF correlate with the human perception of tumor texture shape and therefore, reflect the human impression of an image [27].
To assess the entire baseline tumor burden, 109 metastases (median, 5 VOIs per patient, range, 2-17) were initially segmented. After treatment initiation with vandetanib, 56 metastases (median, 2 VOIs per patient, range 1-13) at follow-up were still available for segmentation. Details on segmented VOI for both baseline and follow-up PET are given in Supplementary Table 2. Change in % between both PETs was calculated using the following formula for every

Tumor marker assessment and clinical parameters
Serum levels of carcinoembryonic antigen (CEA, mg/L) and calcitonin (CTN, pg/ml) were measured (prior to baseline imaging with a median of 6 determinations) [7]. Tumor marker doubling times were calculated using the American Thyroid Association Calculator [28]. Other investigated clinical parameters were metastatic sites at baseline PET, prior therapies, sex, age, and tumoral RET mutation status [21].

Statistical analysis
Statistical analysis was performed using Medcalc (Vers. 17.4.4). The cutoff values of each parameter for PFS and OS prediction were determined by receiver operating characteristic (ROC) analysis (with the Youden-Index for maximization of specificity and sensitivity) (23). Kaplan-Meier analysis (univariate analysis) was performed using thresholds established by ROC analysis in cases in which ROC showed statistically significant results. A multivariate Cox hazard analysis was conducted to determine independent prognostic parameters (24,25). Additionally, Hazard Ratios (HR) were obtained. Statistical significance was considered with a P value < 0.05.

Imaging-based volumetric parameters and textural features derived by 18 F-FDG baseline and follow-up PET
The mean (range) values for all investigated PET parameters are given in Table 1 (for baseline 18  None of the investigated delta parameters reached significance in ROC analysis. Table 2 summarizes the results of all parameters that demonstrated significance in a ROC analysis. Supplementary Table 3 gives an overview of OS and best response for every patient above/beyond the threshold for Complexity and TLG. Of note, 8/10 (80%) of the patients above the ROC-derived threshold for Complexity (for OS) were also in the high-risk as determined by the volumetric parameter TLG.
To determine independent prognosticators, a subsequent multivariate Cox analysis was performed: for PFS, baseline TF Contrast was found to be significant (p = 0.04). For OS, baseline TF Complexity and baseline TLG reached significance (p < 0.0001, respectively).
Using ROC-derived cutoffs, Kaplan-Meier analysis found a significant separation between high-and low-risk groups for baseline TF Contrast (PFS, p = 0.02). For OS, baseline TF Complexity (p = 0.006) and baseline TLG (p = 0.008) reached significance. Respective Kaplan-Meier Plots for selected parameters (all obtained from baseline 18 F-FDG PET) are displayed in Fig. 1.

Discussion
In our previous investigation, SUV mean of a pretherapeutic 18  able to predict PFS for MTC patients undergoing treatment with vandetanib. However, this conventional approach of solely focusing on the metabolically most active lesion failed to reliably predict OS [21]. Hence, in the present study, the entire tumor burden derived by baseline and follow-up PET was re-analyzed. Further, the prognostic potential of tumor heterogeneity was evaluated. Specifically, the higher-order TF Complexity was found to be an independent predictor of patient outcome. Complexity, derived by the neighborhood gray-tone difference matrices, mainly represents a high degree of information content (i.e. an increased number of patches and primitives obtainable in the analyzed texture) [19]. Not surprisingly, Complexity derived by baseline PET portended inferior outcome with significantly reduced OS and trended towards significance for PFS in a ROC analysis (P = 0.07). These findings were further corroborated by obtaining an HR > 6 for cancerrelated death for those subjects above the ROC-derived threshold. Similar results were observed for the volumetric parameter TLG: a higher glucose consumption by the tumor volume at baseline was also independently associated with shortened OS (HR, 9.5; Table 2). Changes in PET features for TKI response assessment and survival prediction have been evaluated in numerous previous studies. However, primarily SUV parameters (SUV peak , SUV max ) have been investigated for outcome prediction, e.g., in NSCLC treated with the TKI erlotinib [29][30][31]. Cook et al. reported on heterogeneity assessment in NSCLC treated with the same TKI and the change of first-order parameters were correlated with survival [19]. However, due to a higher extent of voxel intensities, parameters of first-order intrinsically have a higher risk of association with SUV, particularly in patients with increased SUV max [19,26]. Of note, the range of SUVs derived by the entire tumor burden was rather low in the present investigation (SUV max at baseline, 3.4-19.9 vs. SUV max at baseline in the study of Cook et al., 1.2-30.1 [19]). Consequently, no association between first-order TF and outcome was found in our study. An impact of tumor size on NGTDM-derived TF (such as Complexity) cannot be definitively ruled out [32]: an increase in tumor volume might be associated with an increase in necrosis/hypoxia, which ultimately results in higher informative content within a VOI. However, Complexity was significant in both ROC and Cox analysis for OS, which emphasizes its independent statistical value from other tumor-size related parameters, such as MTV. Notably, discordances between (morphological) responses and outcomes as assessed by RET mutations in MTC demonstrate distinctive mutational heterogeneity and the mutational status can differ between primary tumor and other lesions and even among synchronous or metachronous metastases [33]. Although in the present study, a RET mutation could only be proven in a minority of the cases (3/8, 37.5%), these different oncogenic mechanisms underscore the complexity of tumor biology in progressive MTC. Additionally, vandetanib perse leads to inhibition of tyrosine kinases resulting in reduced tumor proliferation and angiogenesis [34] and novel approaches to further investigate this potential dedifferentiation under treatment are being intensively sought. Several studies have hypothesized that intratumoral heterogeneity derived by PET or CT is linked to angiogenesis or hypoxia [35]: both factors are known to be histopathological correlates denoting either true tumor escape or more aggressive tumor behavior [36]. Although the herein presented strategy of prognostication by using PET-based tumor texture in advanced MTC should be rather interpreted as a "proof-of-concept", it might open avenues to identify high-risk patients at an earlier stage or to monitor those patients more closely.

F-FDG PET derived by a simplistic region of interest was
The herein obtained data must be interpreted with extreme caution: First, although a homogenous cohort has been studied, the number of included patients in the present investigation is rather low, thus limiting statistical power. A prospective trial on a larger scale could further strengthen our preliminary findings. Second, different scanners at different treatment sites were used and therefore, imaging procedures varied from center to center. Thus, an impact on semi-quantification and on the reproducibility of the derived findings cannot be excluded. Apart from that, manual segmentation itself is prone to observer bias; however, experienced nuclear medicine physicians have determined the tumor volume in a consensus analysis. Moreover, the herein presented parameters should be easily obtainable, e.g. by a (semi-)automatic lesion detection software: this might pave the way for a cost-effective and time-saving implementation of TF assessment in clinical routine in the long run [37]. Future efforts may also investigate whether  and to what extent pre-treatment (e.g. by loco-regional procedures) has an impact on intra-tumoral heterogeneity and outcome correlations.

Conclusions
The TF Complexity and the volumetric parameter TLG, obtained from baseline PET, are both independent parameters for OS prediction in MTC patients scheduled for TKI treatment. Further investigations using intra-tumoral heterogeneity for risk stratification and prognostication are warranted, particularly in patients suffering from tumor entities treated with TKI.

Compliance with ethical standards
Conflict of interest RB has a non-commercial research contract with Mediso Medical Imaging Systems, RB is on the speaker's bureau for Mediso Medical Imaging Systems and consultant for Bayer Health-Care. Mediso Medical Imaging Systems employs NZ. All other authors declare that they have no conflict of interest.
Research involving Human Participants All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. For this type of study formal consent is not required. This article does not contain any studies with animals performed by any of the authors.
Informed Consent Informed consent was obtained from all individual participants included in the study.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://crea tivecommons.org/licenses/by/4.0/), which permits use, duplication, adaptation, distribution, and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.