Combinative evaluation of primary tumor and lymph nodes to predict pelvic lymphatic metastasis in cervical cancer: an integrated PET-IVIM MRI study

Background The aim of this study was to evaluate the value of combining pelvic lymph node and tumor characteristics on positron emission tomography-intravoxel incoherent motion magnetic resonance (PET-IVIM MR) imaging for predicting lymph node metastasis in patients with cervical cancer, especially in those with negative lymph nodes on PET. Methods The medical records of 95 patients with cervical cancer who underwent surgical resection with pelvic lymph node dissection were evaluated. The patients were divided into negative and positive groups according to postoperative pathologic lymph node diagnosis, and comparisons of the PET and IVIM-derived parameters between the two groups were performed. Univariate and multivariate analyses were performed to construct a predictive model of lymph node metastasis. Results For all patients, tumor SUVmax, TLG, Dmin, PET and MRI for lymph node diagnosis showed significant differences between patients with and without confirmed lymph node metastasis. Univariate and multivariate logistic analysis showed that the combination of tumor TLG, Dmin and PET for lymph node diagnosis had the strongest predictive value (AUC 0.913, p < 0.001). For patients with PET-negative lymph nodes, SUVmax, SUVmean, MTV, TLG, and Dmin showed significant between-group differences, and univariate and multivariate logistic analysis showed that TLG had the strongest predictive value. Conclusions The combination of tumorTLG, Dmin and PET for lymph node diagnosis is a powerful prognostic factor for all patients. TLG has the best predictive performance in patients with PET negative lymph nodes.


Background
Cervical cancer, one of the common malignant tumors of the female reproductive system, is a serious threat to women's health and life [1]. Local recurrence and distant metastasis are the main causes of death. At present, treatment and prognosis plans for patients with cervical cancer are mainly based on the Federation International of Gynecology and Obstetrics (FIGO) stage. According to FIGO 2018 for cervical cancer, regardless of tumor size and parametrial infiltration, the involvement of lymph node metastasis is classified as stage IIIC. Therefore, accurate diagnosis of lymphatic metastasis is crucial for developing individualized treatment plans, improving prognosis, and reducing mortality [2][3][4].
Recently, fluoro-D-glucose (FDG)-positron emission tomography (PET) has been applied to the diagnosis of cervical cancer metastatic lymph nodes. Rather than computed tomography (CT) or magnetic resonance imaging (MRI), which identifies lymphatic metastasis according to short diameter length and morphology [5].FDG-PET provides quantified metabolic information about lymph nodes and is widely used in the evaluation of lymph node metastasis [6]. One study has proved the prognostic value of lymph node metabolism information in patients with cervical cancer [7]. A lymph node with significantly higher FDG metabolism than the background level is defined as a PET-positive lymph node [8,9]. However, FDG is not a specific imaging agent, and lymph node reactive hyperplasia is also characterized by high metabolism resulting in many false positive cases. Also, the limited resolution of PET and partial volume effect will also affect the diagnostic accuracy of small lymph nodes.
Therefore, quantitative analysis of lymph nodes alone is ambiguous to achieve satisfied accuracy for predicting lymph node metastasis. In this study, we present a combined model (PET and intravoxel incoherent motion (IVIM)-derived imaging of primary tumors and PET/ MRI diagnosis of lymphatic metastasis) for predicting lymphatic metastasis confirmed by postoperative pathology in all patients and patients with PET-negative lymph nodes.

Patients
We retrospectively collected the medical records of patients with a high suspicion of cervical cancer who had underwent PET-IVIM MRI in our hospital from April 2017 to September 2019. The study participants met the following inclusion criteria: (1) underwent surgery; (2) did not receive radiotherapy or chemotherapy before surgery; and (3) had confirmed cervical cancer via postoperative pathology.

Pathologic diagnosis
Postoperatively, the tissue was transported to the pathology department of our hospital. Hematoxylineosin (HE) stained sections were evaluated by two pathologists with more than 10 years of experience. The following information was recorded: lymph node metastasis, histologic tumor type, cervical stromal invasion depth, and tumor differentiation grade. The patient with one or more pelvic metastatic lymph nodes identified with pathological results was regarded as a positive case.

PET-MRI scanning and image acquisition
All patients underwent GE Signa integrated PET/MRI (Signa, GE Healthcare), which combined a 3.0 T MRI scan (GE Signa 750w) and TOF-PET, allowing simultaneous collection of both PET and MRI data. Before the examination, the patients fasted for 4 to 6 h with measured blood glucose levels lower than 7.0 mmol/L. 18 F-FDG (4 MBq/kg) was injected through the cubital vein in the resting state, and PET/MRI was performed after 60 ± 12 min with a 32-channel coil (Upper Anterior Array, UAA). For PET scanning execution, a Dixon MRI sequence was used to attenuate the gamma rays, and LIST-mode and the ordered subset maximum expected iteration method were used to reconstruct the images. The MR-IVIM imaging parameters are summarized in Table 1. PET/MRI acquisition sequence and durations are shown in Fig. 1.

Image analysis
Image analysis using AW4.6 (GE Medical System) workstation was performed by 2 radiologists with more than 4 years of radiodiagnostic experience. After delineation of a lesion region of interest (ROI) at the optimal level of the PET image, the fused-PET/MRI software automatically calculated the metabolic tumor volume (MTV), total lesion glycolysis (TLG), and maximum and mean standardized uptake values (SUV max and SUV mean ) of the entire tumor. A 40% SUV max threshold was used to calculate MTV [10,11]. The IVIM data were analyzed using IMAgenGINE MRToolbox software (Vusion Tech Ltd). The IVIM formula was: Sb/S0 = F exp. [−b × (D* + D)] + (1-F) × exp-(b × D) [12]. On the axial T2-weighted image, the radiologists delineated ROIs at all levels of the lesion, and then the software automatically generated the volume of interest (VOIs) and copied them to the apparent diffusion coefficient (ADC) map, diffusionrelated coefficient (D) map, perfusion related diffusion coefficient (D *) map, and perfusion-related parameter (f) map, obtaining the parameters ADC mean , ADC min , D mean , D min , D * , and f.

PET and MRI diagnosis of lymphatic metastasis
By comparing with surrounding background tissues, a lymph node with increased uptake by visual assessment is considered as a PET-positive lymph node, independently of node size. The method of PET image analysis was based on the previous literatures [8,9]. Patients with one or more positive lymph nodes on PET imaging were considered to be patients with PET positive lymph nodes. Lymph nodes with short axis diameter greater than 1.0 cm and morphological characteristics on MRI were considered as MRI-positive lymph nodes [13,14]. The results were independently evaluated by two radiologists with 12 and 13 years of experience in nuclear medicine and radiology.

Statistical analysis
The data were analyzed using MedCalc software (version 15.2.2), and a p value less than 0.05 was considered as statistical significance. The patients were divided into negative and positive group according to postoperative pathological lymph node diagnosis. Differences in lymph node diagnosis by PET/MRI between groups were evaluated using the chi-square test. The values of PET-and IVIM-derived parameters of the primary tumor were used for the quantitative statistical analyses and the between-groups comparison was performed using the ttest or Mann-Whitney U test. Interclass correlation coefficient (ICC) was used to determine the consistency of the results recorded by the two radiologists (0.8-1 indicated high consistency) [15]. Prediction models for pelvic lymph node metastasis were constructed by univariate and multivariate logistic regression (variables with p < 0.1 in the univariate analysis were used in the multivariate analysis. Bonferroni corrections were performed to correct multivariate analysis). Receiver operating characteristic (ROC) curves were generated for

Patient characteristics
Ninety-five patients were enrolled ( Fig. 2): 78 without lymph node metastasis and 17 with lymph node metastasis (Figs. 3 & 4). The clinicopathological characteristics of the participants are summarized in Table 2.

ICC statistics
Interobserver agreements were excellent for tumor SUV-

The relationship between lymph node metastasis and imaging parameters in all patients
In all patients, SUV max (p = 0.048), TLG (p = 0.014), D min (p = 0.020), MRI in lymph node diagnosis (p = 0.009) and PET in lymph node diagnosis (p < 0.001) showed significant differences between patients with and without lymph node metastasis ( Table 3). The remaining PETand IVIM-derived parameters did not show statistical differences between the two groups. ROC analysis of the negative and positive groups according to postoperative lymph node diagnosis showed that SUV max (AUC 0.654, p = 0.023), SUV mean (AUC 0.646, p = 0.030), TLG (AUC 0.692, p = 0.009), D min (AUC 0.681, p = 0.007), MRI in lymph node diagnosis (AUC 0.656, p = 0.020) and PET in lymph node diagnosis (AUC 0.736, p < 0.001) had a positive effect on predicting metastatic lymph nodes confirmed by postoperative pathology (Fig. 4). The optimal cut-off threshold values for SUV max , SUV mean , TLG, and D min were 12.98 g/cm 3   Univariate logistic analysis showed that MTV (p = 0.014), TLG (p = 0.019), D min (p = 0.026), MRI for lymph node diagnosis (p = 0.012) and PET for lymph node diagnosis (p < 0.001) were associated with lymph node metastasis. Bonferroni corrected alpha value of 0.005 per test (0.05/10) was applied to the final multivariate analyses in exploring lymph node metastasis. The combination of TLG, D min , and PET for lymph node diagnosis had the strongest predictive value on multivariate logistic analysis ( Table 4). The area under the ROC curve for the combination of TLG, D min , and PET for lymph node diagnosis (AUC 0.913, p < 0.001) was higher than that of any individual parameter (p < 0.05) (Fig. 5). The correlations between parameters are shown in Table 5.
The relationship between lymph node metastasis and imaging parameters in patients with PET negative lymph nodes  (Table 6).

Discussion
Recognition of lympha node metastasis before treatment in patients with cervical cancer is essential for personalized treatment plans, and diagnostic imaging is most commonly used for this purpose. In the present study, all data were measured by integrated PET-MRI. With the integrated device, a shortened scanning time, a reduced radiation dose, and the simultaneous PET and MRI parameters can be achieved, enabling combined multiparameter prediction and correlation [16][17][18][19].
The IVIM sequences was introduced into the present PET/MR study, which is an extension of diffusionweighted imaging (DWI) sequence, to PET/MRI. DWI reflects the diffusion dynamics of tissue water molecules by ADC; however, the attenuation of DWI signals in tissues is determined by both water molecule diffusion and microcirculation perfusion. ADC is a quantitative parameter based on single index model of DWI, and it reflects both the true diffusion of water molecules and the "false diffusion", caused by the blood microcirculation in the capillaries. This inability to reflect the movement of water molecules in living tissues limits the utility of DWI to evaluate microstructure changes.
The IVIM model compensates for these technical defects of traditional DWI. The theoretical basis is to doubleexponentially fit tissue signals based on multiple b-values, and quantitatively extract the main signal attenuation components from the low b-values of the curve. The true diffusion parameter value of the tissue was calculated at high bvalues [20]. So, this model has a higher accuracy to reflect the signal attenuation of water molecules than the DWI single exponential model. IVIM models can calculate three parameters, including D (purediffusion coefficient), f (perfusion fraction), and D* (pseudo-diffusion coefficient). In the IVIM model, D and f value are indicators to reflect the true diffusion coefficient and perfusion information of the molecule, respectively. IVIM modalities have also been used in other cervical cancer studies [21,22]. However, the current IVIM technology still has shortcomings and has not been applied on a large scale in the clinic setting. The limitations include: (1) The organs in the abdominal cavity may have a slight displacement movement during long-time scans, which may affect image post-processing; (2) image distortion caused by magnetic susceptibility artifacts at the gassoft tissue interface leads to artifacts obscuring lesions or unclear display of lesions, affecting data measurement and (3) imprecise ROIs including normal tissues result in poor measurement repeatability.
We established a combined prediction model for cervical cancer patients with lymphatic metastasis by multiple logistic regression analysis, using TLG and D min of the tumor and PET for lymph node diagnosis while filtering out the remaining parameters. Several mechanisms may explain the strong predictive potential of these 3 combined parameters. In previous studies, lymph node with FDG uptake more than the background lymph node metabolism level was a diagnostic criterion for lymphatic metastasis [8,9]. Our study confirms the ability to diagnose lymphatic metastasis with this method (AUC 0.736, p < 0.001), but also demonstrates its low sensitivity (58.82). In contrast, the combination of tumor TLG and D min and this diagnostic criterion had better predictive value than the diagnostic criterion alone (AUC 0.913, sensitivity 94.12, specificity 82.05, p < 0.001; difference between areas 0.176, Z = 3.141, p = 0.002). Previous studies have shown that the expression of glucose transporter-1 (Glut-1) is related to lymph node metastasis in a variety of tumors [23,24], which will trigger an increase in FDG uptake and the corresponding quantitative indicators [25]. TLG is a comprehensive parameter that reflects the metabolic activity of the whole tumor. Compared with SUV max , TLG can  [26]. Similar conclusions have been obtained in studies involving endometrial carcinoma patients and thyroid carcinoma patients [27,28].
Recently, many studies have quantitatively analyzed lymph nodes by IVIM to predict pelvic lymph node metastasis [29,30]. However, the slice of IVIM imaging is thick, which makes it difficult to carry out clinically concerning the usually small volume of lymph nodes. In this study, we evaluated lymphatic metastasis using IVIM parameters of tumors for cervical cancer such as the D value. Theoretically, the D value is not affected by the microcirculation blood flow of the tumor tissue, giving a more accurate signal on the diffusion of the tissue water molecules than does the ADC value. Studies have suggested that D values correlate with hypoxia and the tumor-stroma ratios [18,31]. D min value corresponds to the lowest tumor diffusion region, which may more accurately reflect the number and heterogeneity of tumor cells, as well as tumors with lymphatic metastasis.
A study using similar methods suggests that a combination of serum squamous cell carcinoma antigen (SCC-Ag) level, SUV max , and lymph node status is an important prognostic indicator for cervical cancer [32]. Another study suggested that the combination of lymph node CT indicators and tumor TLG values can improve the accuracy of PET/CT for the diagnosis of lymphatic metastasis in patients with cervical cancer [33]. Similar combined methods are also applied to breast cancer [34].
A previous study has shown that PET/MR for lymphatic metastasis was mainly dependent on the sensitivity of PET [35]. In this study, PET was also significantly better than MR in the diagnosis of lymph nodes. But we found a large number of PET false-negative cases, which accords with the findings of endometrial cancer and non-small cell lung cancer studies [36,37]. Identifying PET false-negative cases has always been a clinical challenge. Therefore, we further explored the relationship between lymph node metastasis and imaging parameters in patients with PET/MRI-negative lymph nodes. By multivariate logistic analysis, TLG had the strongest predictive value for lymph node metastasis in this subset of patients.
This study has several limitations. On one hand, our scan range was from the vaginal level to the upper edge of the humerus. Therefore, we could not measure the SUV peak value which may better reflect the aggressiveness of the tumor [25]. On the other hand, we used the pathologic diagnosis as the gold standard. The inclusion of patients who have undergone surgery results in fewer negative cases in the sampled population. A further evaluation of more cases is needed to verify our conclusions.

Conclusions
In all patients, the combination of tumor TLG and D min of the tumor and PET for lymph node diagnosis had better predictive performance than the other imaging parameters for nodal metastasis. In patients with PET negative lymph nodes, TLG showed the strongest predictive potential. These findings may have clinical applications in personalized treatment planning for patients with cervical cancer.