In this study, a nomogram based on logistic regression and a machine learning model exploring RF method was established for the differential diagnosis of HP and GIST. Among significant independent predictors, age, location, LD/SD ratio and HU lesion / HU pancreas A were shared by the two models, except for the unique duct-like structure in the nomogram and peritumoral hypodensity line in the RF model. The results confirmed the two models’ good discrimination and calibration, but the nomogram might result in a larger net benefit than the RF model according to decision curve analysis.
HP, which is defined as a kind of ectopic flat glandular tissue and shares similar histological composition with normal pancreas, tends to manifest as broad-based lesions with wall-attached growth pattern, and thus often has a larger LD/SD ratio. But GIST is a true tumor that prefer to grow towards smaller ratio. LD/SD ratio was an important diagnostic predictor in two models with median value of 1.50 and cut-off point of 1.47, which is essentially consistent with many previous studies [2, 4, 22, 30].
As for location, there was no statistically significant difference in the distribution of HP and GIST on the stomach and small intestine. The result of HP is similar to a clinicopathological study [31], in which the stomach (97/184, 52.7%) was the most common location in the gastrointestinal tract, followed by the small intestine (48/184, 26%). However, presumably owing to sample size error, the incidence in the small intestine was significantly higher than in the stomach in other studies, particularly in the duodenum and proximal jejunum [13, 32]. In particular, GISTs in our series were more often located in the upper part of stomach (49, 49%), followed by the lower part of stomach (17, 17%), whereas most HPs were located in the lower part of stomach (25, 52.1%), followed by the duodenum (7, 14.6%). This distribution tendency of HP is consistent with many studies [2, 6, 20, 23, 31], and can be explained by “misplacement hypothesis”—HPs are fragments split from the main pancreas during embryonic rotation [30, 33].
Since HP often shows bright contrast enhancement similar to that of the main pancreas in arterial and/or venous phase [4, 20, 34], we calculated enhancement ratio of lesion to pancreas in both phases. Finally, some semi-quantitative and quantitative parameters showed significant differences in the degree of reinforcement between HP and GIST—HU lesion / HU pancreas A, HU arterial, DEAP, and enhancement grade, from which we inferred that HP might be more characteristically hyperenhanced in the arterial phase. Only HU lesion / HU pancreas A was included in both models as an independent predictor of HP, and when the ratio is greater than 0.805, the lesion is more likely to be diagnosed as HP. Meanwhile, we also deduce that the ratio of HPs to pancreas in portal venous phase was closer to 1 (Table 2, median values of HP and GIST were 1.03, 0.90 respectively), which requires more studies to validate further.
Consisted of pancreatic acini, ductal components at different proportions, HP has three subtypes in histologic specimens—acini-dominant, duct-dominant and mixed type [9]. Based on above typing, we further classified the lesions into three types on imaging—solid-dominant, cystic-dominant and mixed type. Li et al. [20] removed the complete cystic HPs, in fact there were also 2 cases of cystic GIST in our study, and no statistical differences was found in image type between the two lesions finally. Other CT features, including ill-defined border, microlobulated appearance [30], presence of low intralesional attenuation and enhancement pattern [22, 30], were also associated with lobular architecture or dilated residual duct of HP [2, 4, 6], and were deduced by prior studies to be distinctive signs of HP. Yet they had not statistically significant difference or were just relevant predictors in our series. Due to HP likely extending to the muscularis propria or the entire wall of gastrointestinal tract [31, 35], we propose “peritumoral hypodensity line” to describe the relationship between the lesion and the peripheral gastrointestinal wall, which could reveal lesions as clear submucosal lesions or represented fat space between the extraluminal lesion and serosal layer. It was included in the RF model finally.
Duct-like structure was found to be a significant CT feature with secondary importance in the nomogram, which was referred to a central umbilication located at the mucosa of the lesion, corresponding to the rudimentary duct of the HP as seen in histologic specimens [30, 34]. T2-weighted MR images and magnetic resonance cholangiopancreatographic (MRCP) (Fig. 2B) images are best for confirming a dilated duct in HP, which is also referred to as the “ectopic duct” sign [4, 8, 11]. This sign was not included in the RF model. We infer that this sign, while unique to the ectopic pancreas, was relatively difficult to be observed. Only 7 of 48 HPs were detected this CT morphologic feature in this study, with 2 located in the lower part of stomach, 3 in the duodenum, and 2 in the jejunum or ileum, respectively. We inferred from this that the sign was easier to discern in the small intestine, similar to some previous reports [5, 30]. In addition, the trend in the age distribution of HP and GIST is consistent with most of the studies we've seen.
Our study has several limitations. First, the retrospective design may have introduced inherent selection bias, although our patients were enrolled from two institutions. Second, due to the long period of follow-up, different CT machines and protocols were used, which might influence the quantitative analysis. Third, due to the use of lesions LD less than 3 cm in the inclusion criterion, we excluded a great proportion of GISTs. Fourth, the size of our study population was small, especially for HP, and thus the stability of the model may be affected and the false-positive rates may increase.