Pancreatic neuroendocrine tumors (PNETs) are rare tumors of the pancreas with an increasing incidence that can be attributed to increased utilization of cross-sectional imaging [1], [2], [3]. PNETs are heterogeneous in terms of histopathological characteristics and clinical behavior, ranging from indolent to very aggressive [4,5]. Identification of factors that are indicative of tumor behavior is essential to determine appropriate management for each patient. One of the most useful factors in predicting outcome is tumor size. However, it has been shown that a subset of small tumors (< 2 cm) can exhibit aggressive behavior. Histological grade, as determined by the proliferation rate of the neoplastic cells is therefore also critical in predicting the behavior of PNETs.

World Health Organization (WHO) grading scheme (referred subsequently to as tumor grade) was revised in 2017 and 2019, and it currently uses the proliferation rate of the neoplastic cells, as determined either by mitotic rate or Ki-67 labeling index, to stratify PNETs in to G1, G2, G3 tumors [6], [7], [8], [9]. Even though the WHO scheme was developed based on the complete assessment of surgically resected specimens, the scheme has been applied to limited tissue obtained via endoscopic ultrasound and fine needle aspiration (EUS-FNA) [10], [11], [12]. Recently, it has been shown that EUS-FNA can accurately predict tumor grade in small PNETs, however, for larger tumors given their intratumoral heterogeneity a lower accuracy has been reported [13]. Furthermore, sampling the tumor for grading requires invasive procedure, and for patients on surveillance serial assessment of changes in tumor grade is not feasible. Therefore, there is a need to develop novel tools for non-invasive assessment of tumor grade in PNETs.

Recently the field of radiomics has burgeoned and has shown promise as a potential tool for tumor characterization and patient prognostication. Radiomic analyses extract high-dimensional data from images and provide feature data for quantitative description of lesions [14]. This approach has demonstrated good performance in differentiating between tumor types, estimating tumor grade, assessing treatment response, and predicting survival for multiple tumor types [15], [16], [17], [18], [19], [20], [21]. Literature on the application of radiomics to assessment of tumor grade in PNETs is limited, with a majority of studies evaluating relatively small patient populations [22], [23], [24], [25].

The goal of this study was to develop a predictive model using radiomic features to predict tumor grade in PNETs and compare its performance to that of EUS-FNA.