ABSTRACT

PURPOSE

The purpose of this paper was to distinguish solid pseudopapillary neoplasms (SPNs) and nonfunctional neuroendocrine tumors (nf-NETs) of pancreas using univariate analysis and clinical-CT logistic regression model.

METHODS

Twenty-eight patients with SPNs and 46 patients with nf-NETs underwent enhanced CT examinations. Clinical data (sex, age), categorical (location, cystic degeneration, calcification, hemorrhage, and enhancement pattern), and numeric CT features (lesion long diameter, long/short diameter ratio, tumor attenuation values and tumor/pancreas attenuation ratios at unenhanced phase [UP], arterial phase [AP], and venous phase [VP]) were recorded. The logistic regression model was constructed by stepwise forward method of binary logistic regression after univariate analysis. The corresponding operating characteristic curve (ROC) and nomogram were delineated. The area under the curve (AUC), sensitivity, and specificity of ROC were calculated.

RESULTS

The SPNs were observed more often in relatively young (P < .001), female (P < .001) patients. After the univariate analysis, the categorical CT features of location (P = .048), hemorrhage (P = .003), and enhancement pattern (P = .004) and the numeric CT features of lesion long diameter (P = .005), tumor/pancreas_UP (P = .002), tumor_AP (P < .001), and tumor/pancreas_AP (P < .001) had statistical significance. The AUC (95% CI), sensitivity, and specificity of a logistic regression model composed of age, tumor/pancreas_UP, and tumor/pancreas_AP were 0.933 (95% CI, 0.850-0.978), 84.78%, and 92.86%.

CONCLUSION

The SPNs often occurred in 20- to 40-year-old female patients, were located in the body or tail of pancreas, showed hemorrhagic degeneration, heterogeneous enhancement, and were relatively larger in size compared with nf-NETs. Tumor/pancreas_UP, tumor_AP, and tumor/pancreas_AP values of SPNs were smaller than those of nf-NETs. The clinical-CT logistic regression model and nomogram consisting of age, tumor/pancreas_UP, and tumor/pancreas_AP parameters helped to differentiate SPNs from nf-NETs.

Solid pseudopapillary neoplasm (SPN) of pancreas is an uncommon and unique entity accounting for 1%-3% of all pancreatic tumors,¹ which was initially reported by Frantz in 1959² and designated as SPN by the World Health Organization (WHO) in 1996.³ According to the latest WHO classification of digestive tumors,⁴ SPN presents a low malignant potential and excellent long-term prognosis following surgical resection. Pancreatic neuroendocrine tumor (NET) is a neoplasm derived from progenitor islet cells⁵ and has been divided into functional NET (f-NET) and non-functional NET (nf-NET) based on their clinical symptoms caused by different hormones⁶ and the hormone levels detected by blood tests.

SPNs have various pathological and radiological features that overlap with other pancreatic neoplasms, in particular, nf-NETs. Microscopically, both SPNs and NETs comprise a monotonous and solid cellnest.⁷ Similar pseudopapillary-like cellular arrangement and expressionof CD56and synaptophysin further complicate the differentiation, while the diffused nuclear expression of βcaptein is of great significance for the diagnosis of SPNs.⁷ Typically, most SPNs of the pancreas are large, well-encapsulated, and heterogeneous masses with areas of hemorrhagic, cystic, or calcific degeneration⁷ on computed tomography (CT) or magnetic resonance imaging (MRI). However, nf-NETs have a number of atypical characteristics, mimicking the presentation of SPNs on CT or MRI images. Thus, nfNETs should be at the top of the differential diagnosis list. SPNs have a significantly better prognosis and survival compared to nf-NETs. Therefore, the distinction between them is of paramount importance.

To the best of our knowledge, few categorical and numeric CT logistic regression analyses have been reported to differentiate SPNs from nf-NETs. This retrospective study attempted to construct a clinical-CT logistic regression model to better distinguish these two diseases by systematic CT analysis.

Methods

This retrospective study was approved by our institutional review board (No. 2020 QT338) and the formal written consent was waived.

Patients

We reviewed all the pancreatic data at the picture archiving and communication system (PACS) from January 2013 to September 2020. The inclusion criteria were as follows: (a) all the patients were pathologically proven to have either SPNs or nf-NETs after surgical resections, (b) patients underwent enhanced CT examinations with the same scan protocol, (c) patients received preoperative CT examinations within 3 months before surgeries, and (d) the nfNET patients did not have clinical symptoms caused by secreted hormones and had normal hormone levels after blood tests. The exclusion criteria were as follows: (a) patients got radiation or chemotherapy treatment before CT examinations; (b) the quality of CT images were not good enough for analysis because of motion or breathe artifacts, parts of images lost, and so on; and (c) patients accompanied by other pancreatic diseases, such as inflammation or tumor. The specific workflow of patient recruitment was shown in Figure 1.

CT examinations

All three-phase enhanced CT scans were performed using a 64/128-multidetector CT (SOMATOM Definition AS, Siemens Healthcare). The three phases of CT examination consisted of unenhanced phase (UP), arterial phase (AP), and venous phase (VP). Using the method of computeraided bolus tracking technology and setting a 100 HU in the abdominal aorta as the threshold baseline, the AP and VP scans were performed with approximately 30 s and 60 s delay after administration of 90-100 mL nonionic contrast material (lopromide, 370) via an antecubital vein at a rate of 3.0 mL/s. The tube voltage was 120 kVp and the tube current was 200 mA. The collimation was 64 × 0.625 mm and rotation time was 0.75 s. The slice thickness and intervals of images were both 5.0 mm.

CT image analysis

The CT image analysis included categorical and numeric assessments. Two radiologists (7 years and 10 years experience in abdominal diagnosis) worked together to analyze categorical CT images in consensus and blinded to the pathological results. The categorical features included location (head, body, or tail of pancreas), cystic degeneration (defined as the area with CT attenuation between that of fluid and soft tissue⁸ (Figure 2a, 2d), calcification (Figure 2b, 2c), hemorrhage defined according to the high-attenuation component greater than 45 Hu presented in the lesion⁹ (Figure 2b) and enhancement pattern categorized as homogeneous, mildly heterogeneous, and markedly heterogeneous by visual evaluation¹⁰ (Figure 2c, 2d).

The numeric characteristics were measured by two radiologists, respectively, and the mean values were taken for further analysis. The intraclass correlation coefficient (ICC) was calculated to assess the agreement between two radiologists. The regions of interest (ROIs) of tumor and pancreas were measured in the longer axis of the lesion and included the solid component of the lesion as much as possible.¹¹ The numeric characteristics were as follows: long diameter of the lesion, long-toshort diameter ratio (long/short diameter ratio), tumor CT attenuation on UP (tumor_UP), tumor-to-pancreas CT attenuation ratio on UP (tumor/pancreas_UP), tumor-enhanced CT attenuation on AP (tumor_AP, calculated as the difference value of tumor CT attenuation between the AP and UP), tumor-to-pancreas CT attenuation ratio on AP (tumor/pancreas_AP), tumor-enhanced CT attenuation on VP (tumor_VP, calculated as the difference value of tumor CT attenuation between the VP and AP), and tumor-to-pancreas CT attenuation ratio on VP (tumor/pancreas_VP).

Statistical analysis

The analysis of categorical CT features was analyzed using Pearson chi-square test or Fisher exact probabilities method. The analysis of numeric CT features was analyzed by method of independentsample t test or Mann–Whitney U-test if variables distributed non-normally. The skewness and kurtosis values were calculated to test the normality of numeric variables. The normal distributed variables were described as average ± standard deviation, while the non-normal distributed ones were described as median (min-max) values. Then, the significant risk factors were enrolled to construct logistic regression model by the method of stepwise forward binary logistic regression. The HosmerLemeshow test was made to evaluate the fitting degree of the logistic regression model. These statistical methods were performed using IBM SPSS Statistics (Version 22.0). The ICC of numeric CT features between two radiologists and the ROC of corresponding logistic regression model were made on software of MedCalc (Version 18.2.1). The relative nomogram was depicted using R software (Version 4.1.0). A two-tailed P < .05 was considered statistically significant. The P value > .05 of HosmerLemeshow test indicated the goodness of fit. The ICC > 0.75 suggested good consistency between different observers.

Results

Results of clinical data and CT characteristics were summarized in Table 1. Typically, SPNs occurred more frequently in female (P < .001) and younger patients than nf-NETs (34.14 ± 13.47 vs. 53.80 ± 13.23, P < .001).

In categorical CT analysis, cystic generation and calcification showed no statistical significance between SPNs and nf-NETs. Meanwhile, hemorrhagic (P = .003), heterogeneously enhanced (P = .004) masses in body or tail (P = .048) of pancreas tended to be diagnosed as SPNs.

In numeric analysis, the ICC between radiologists was 0.782-0.895, which indicated good agreement (ICC > 0.75) between two radiologists. The long diameter (P = .005) of the SPN was generally longer than that of the nf-NET (average ± standard deviation, 55.39 ± 29.46 mm vs. 37.20 ± 22.98 mm). However, the long/short diameter ratio showed no significant difference between the two diseases. Compared with nf-NETs, the values of tumor/pancreas_UP (P = .002, Figure 3a), tumor_AP (P < .001), and tumor/ pancreas_AP (P < .001, Figure 3b) of SPNs were relatively smaller. On the other hand, the tumor_UP, tumor_VP, and tumor/pancreas_VP (Figure 3c) were not statistically different between SPNs and nf-NETs.

Significant risk factors of sex, age, location, hemorrhage, enhancement pattern, lesion long diameter, tumor/pancreas_UP, tumor_AP, and tumor/pancreas_AP were included in the logistic regression analysis. Finally, the clinical-CT logistic regression model (Table 2) was constructed based on three independent predictors of age (P = .005), tumor/pancreas_UP (P = .034), and tumor/pancreas_AP (P = .004). The nonsignificant Hosmer-Lemeshow (P = .674) showed suggested goodness of fit of this logistic regression model. The equation of the model was: Y = − 10.232 + 0.082* [age] + 5.809*[tumor/pancreas_UP] + 9.146* [tumor/pancreas_AP]. Then, the ROC (Figure 4) and nomogram (Figure 5) were delineated according to this logistic regression model. The AUC (95% CI), sensitivity, and specificity of ROC in distinguishing SPNs and nf-NETs were 0.933 (95% CI, 0.850-0.978), 84.78%, and 92.86%.

Discussion

To date, only a few publications compared the imaging features between SPNs and NETs using univariate analysis,¹² and there has been no CT-based logistic differentiation of SPNs and nf-NETs. Among the 28 SPNs patients enrolled in this study, 7 patients (25.0%) were misdiagnosed as nf-NETs. Meanwhile, 14 out of 46 nf-NETs patients were misdiagnosed, of which 6 (13.0%) were misdiagnosed as SPNs. Thus, the distinction between SPNs and nf-NETs is clinically important. The first section of this study focused on the univariate analysis of clinical and CT features. The SPNs usually showed the following categorical features: female predominance (85.7%), younger age (34.14 ± 13.47 years vs. 53.8 ± 13.23 years), location in pancreatic body or tail (78.6%), hemorrhage (17.9%), and heterogeneous enhancement (85.7%). There is a consensus that SPNs tend to occur in young female patients.^{13, 14} Data from previous studies have suggested that SPNs often located in the body or tail of pancreas,¹⁵ also consistent with our result. Ventriglia et al.¹⁶ highlighted that specific MRI features of internal hemorrhage and well demarcation helped to differentiate SPNs from other cystic pancreatic tumors. Intratumoral hemorrhage must be considered as an important pathological characteristics of SPNs¹⁷ even if it was just found in 5/28 (17.9%) in this study. It has been demonstrated that the presence of cystic pattern or hemorrhage in varying proportions are the results of a process of degeneration.¹⁸ However, no statistical difference was noted in features of cystic degeneration and calcification between SPNs and nf-NETs in our study. Heterogeneous enhancement corresponded with tumoral degeneration, which has been histopathologically proved.¹⁹

A previous publication assessing the enhancement manifestation by visual evaluation of MRI illustrated that all SPNs had early heterogeneous and progressive enhancement pattern, but 67% of NETs showed persistent homogeneous enhan cement.²⁰ This paper gave a deta iled measurement of CT attenuation and found that smaller tumor/pancreas_UP, tumor_AP, and tumor/pancreas_AP were associated with SPNs. The relative CT attenuation of tumor/pancreas_UP, tumor/pan creas_AP, and tumor/pancreas_VP were calculated after standardization against pancreas, reducing the inter-individual variability. Regarding the VP of CT examination, SPNs had similar enhancement with nf-NETs in this analysis, which was not tally with previous results.²¹

After the univariate analysis of clinical data and CT features, a logistic regression model was built, achieving a promising performance in differentiating SPNs from nf-NETs. The AUC of this model was 0.933. This clinical-CT logistic regression model included age, tumor/pancreas_UP, and tumor/pancreasAP. Compared with CT attenuation, the relative attenuation of tumor/pancreas_UP and tumor/pancreas_AP played an important role in differentiating SPNs from nf-NETs, while the VP enhancement did not help to diagnose the diseases significantly, although previous studies had reported that nf-NETs presented a rapid and homogeneous enhancement in MRI images.²¹ This result presumably related to the delay time in VP, though there is no standard scan protocol. The corresponding nomogram achieved straightforward and visualized noninvasive distinction between the SPNs and nf-NETs. This information presented integrated clinical data and CT characteristics and provided a scoring system to facilitate differential diagnosis between two diseases.

There were some limitations in our study. First, the sample size was small because of the rarity of tumors. Moreover, this small sample size restricted the study in morphology of calcification, and different grades of nf-NETs. Second, the characteristics of a tumor capsule of the SPN and floating cloud sign were not analyzed in this paper which need further analysis. Previous studies confirmed that the welldefined capsule facilitated the diagnosis of SPN in MRI images²⁰ and proved focally defective capsule histopathologically.⁷ Third, this retrospective study only differentiated SPNs from nf-NETs because of considerable overlap in CT images. Nevertheless, other pancreatic lesions, such as carcinoma or cystadenoma of pancreas also need future analysis. Fourth, on the numeric analysis of CT attenuation, only the solid portion of the mass was measured and the regions of cystic degeneration, calcification, and hemorrhage were ignored.

In conclusion, the SPNs often occurred in 20- to 40-year-old female patients, were located in the body or tail of pancreas, showed hemorrhagic degeneration and heterogeneous enhancement, and were relatively larger compared with nf-NETs. The clinical-CT logistic regression model and nomogram including age, tumor/pancreas_UP, and tumor/pancreas_AP achieved promising discrimination between SPNs and nf-NETs.