Prognostic Assessment of Gastropancreatic Neuroendocrine Neoplasm: Prospects and Limits of Radiomics

Abstract

Neuroendocrine neoplasms (NENs) are a group of lesions originating from cells of the diffuse neuroendocrine system. NENs may involve different sites, including the gastrointestinal tract (GEP-NENs). The incidence and prevalence of GEP-NENs has been constantly rising thanks to the increased diagnostic power of imaging and immuno–histochemistry. Despite the plethora of biochemical markers and imaging techniques, the prognosis and therapeutic choice in GEP-NENs still represents a challenge, mainly due to the great heterogeneity in terms of tumor lesions and clinical behavior. The concept that biomedical images contain information about tissue heterogeneity and pathological processes invisible to the human eye is now well established. From this substrate comes the idea of radiomics. Computational analysis has achieved promising results in several oncological settings, and the use of radiomics in different types of GEP-NENs is growing in the field of research, yet with conflicting results. The aim of this narrative review is to provide a comprehensive update on the role of radiomics on GEP-NEN management, focusing on the main clinical aspects analyzed by most existing reports: predicting tumor grade, distinguishing NET from other tumors, and prognosis assessment.

1. Introduction

Neuroendocrine neoplasms (NENs) are tumors with significant heterogeneity and complex clinical behavior that originate from cells of the diffuse neuroendocrine system [1].

NENs could be divided into two categories: neuroendocrine tumors (NETs) and neuroendocrine carcinomas (NECs). NETs are well-differentiated, slow growing, but potentially malignant lesions, while NECs are poorly differentiated and highly aggressive carcinomas [2]. NENs may involve many different sites, and in two thirds of cases, may arise in the gastro–entero–pancreatic tract (GEP). Of these, nearly 50% are intestinal (Figure 1) and 30% pancreatic [1] (Figure 2). Among all malignancies in this district, NENs are considered rare (1.0–1.5% of all GEP tumors) [3,4], but their incidence and prevalence has been constantly rising (currently estimated at 3.0–5.2 cases per 100,000 people per year, and 35 cases per 100,000 people per year, respectively) [5] thanks to the increased power of detection through imaging and immuno–histochemistry.

The majority of GEP-NENs (>95%) occurs in sporadic forms, while 5% of cases are part of a polydistrict syndrome, such as multiple endocrine neoplasm type 1 (MEN1), neurofibromatosis type 1 (NF1), and von Hippel–Lindau syndrome (VHL) [6,7]. GEP-NENs could release metabolically active hormones and amines, leading to hypersecretion-specific clinical signs and symptoms, which could facilitate the diagnosis. In this scenario, the diagnosis of functional lesions is clinical, and diagnostic tools are employed to localize the lesion and to stage. Unfortunately, intestinal NENs secrete hormones in only 20%, and pancreatic NENs are functional in only 10–30% [6,7]. So, in these indolent lesions, the presence of non-specific symptoms can cause a delayed diagnosis, with lesions identified in an advanced stage, often for complications due to the primary tumor (bowel obstruction, mass effect) or metastases (especially in the liver) [6,7]. Consequently, an accurate diagnosis requires a multidisciplinary assessment, including a morphological and functional imaging evaluation.

The main prognostic factors for GEP-NETs are the primary site, with a poorer prognosis in the pancreatic location (PNET) (Figure 3); the TNM stage; and the histopathological classification according to the World Health Organization (WHO), which includes the morphological and proliferative criteria based on the Ki67 index [1]. An additional positive prognostic factor is the expression of the somatostatin receptor, which is the basis of imaging diagnosis on Gallium-68 DOTA-TATE Positron Emission Tomography/Computed Tomography (PET/CT) and somatostatin analog (SSA) therapy, including octreotide acetate and peptide receptor radionuclide therapy (PRRT) (such as lutetium Lu 177 dotatate) [6,7].

The prognosis and therapeutic choice in NETs still represents a challenge, mainly due to the great heterogeneity of this family of tumors with varied clinical behavior [8,9].

Regarding imaging evaluation, even if the limits of conventional morphological studies are well recognized, technique optimization, such as technological improvements, with the use of new sequences in magnetic resonance imaging (MRI) have allowed for obtaining a high sensitivity in detection rate. Additionally, conventional imaging allows local staging in order to identify patients fit for surgery, considering that surgical resection, according to ESMO guidelines, is the front line of treatment for many lesions.

According to the guidelines of the European Neuroendocrine Tumor Society (ENETS), contrast-enhanced neck–thorax–abdomen and pelvis CT, which should include a three-phase liver assessment, is suggested during detection, stage, treatment response evaluation, and surveillance phases. Contrast-enhanced MRI, including diffusion-weighted imaging (DWI), is suggested for hepatic, pancreatic, head, and bone evaluation. Ultrasonography (US) with contrast-enhanced ultrasound (CEUS) is suggested as a solution in liver lesions and is a useful tool during surgical resection. In addition, for jejunum and ileum tumors, CT enteroclysis is the diagnostic tool that should be chosen, while in liver and pancreatic lesion assessment, MRI should be chosen. CT permits a vascular assessment in the pre-surgical phase.

With regard to functional evaluation, 68Ga/64Cu-DOTA-somatostatin analogue PET-CT showed the higher sensitivity in NEN detection, and according to ENETS, this tool should be chosen to localize the disease in non-insulinoma pancreatic lesions. In addition, the European Society of Medical Oncology (ESMO) guidelines suggest the use of PET with [18F] fluoro–deoxy–glucose (FDG) as an optional tool in NEN assessment. However, this tool is useful for G3 and high G2 lesions, considering the higher glucose metabolism and the lesser Somatostatin receptor (SSTR) expression compared to low-grade lesions.

The concept that biomedical images contain information about tissue heterogeneity and complex pathological processes is now well established [10,11]. From this substrate comes the idea of radiomics, a quantitative approach to medical imaging that aims to decode the interrelationships between pixels that are invisible to the human eye. From each voxel of a selected tissue region on different imaging techniques, hundreds of features could be extracted and analyzed with multiple post-processing methods and software [12,13]. The ultimate goal is to create faster and more reliable support systems for assisting clinical decision-making and to improve the quality of care [12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37]. The radiomics process includes image acquisition and segmentation, feature extraction and qualification, feature selection, analysis, and classification. The image data are provided by radiological modalities, while feature segmentation consists of the isolation of volume interest regarding the target zone to analyze.

The features were extracted (feature extraction) by means of several tools, such as the Pyradiomics tool 3D Slicer platform (https://pyradiomics.readthedocs.io/en/latest (accessed on 1 January 2021)).

Radiomic features can be divided into six groups: size and shape-based features, image intensity histogram descriptors, image–voxel relationship descriptors, size zone matrix (SZM)-derived plots, and neighborhood grayscale difference matrix (NGTDM), and filtered image textures and fractal features. Two-dimensional and three-dimensional feature extraction can be used to calculate individual feature values for a region of interest (“segment-based”) or to generate feature maps (“voxel-based”). Feature extraction should be carried out according to the Image Biomarker Standardization Initiative (IBSI), an independent international collaboration working to standardize feature extraction.

Due to its huge variety, feature reductions need to be implemented to eliminate redundant information, unstable and unreproducible features that could lead to spurious results and unrepeatable patterns. After selecting the important features, it is essential to analyze the chosen data by first comparing the features with each other to find out if they have any information in common and to reveal what it means when they all occur at the same time. Other methods of testing are supervised or unsupervised testing. The algorithm needs to recognize the correlations between images and features.

Several steps are required to build an integrated radiomics database. Imaging data should be exported from clinics, which is already complex because patient information is very sensitive and governed by privacy laws. Furthermore, the exported data must not lose their integrity when compressed so that the database only incorporates data of the same quality. Furthermore, the integration of clinical and molecular data is important for various clinical outcomes. A large image storage location is required in this case.

Computational analysis has achieved promising results in several oncological settings [38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58], and the use of radiomics in different types of GEP-NET is growing in this field of research [18,59,60], yet with conflicting results.

The aim of this narrative review is to provide a comprehensive update on the role of radiomics on GEP-NETs management, focusing on the main clinical aspects analyzed by most existing reports: predicting tumor grade in PNETs, distinguishing PNET from other tumors, and prognosis assessment.

2. Tumor Classification

The differential diagnosis among types of pancreatic cancers as well as benign diseases could be difficult [61,62,63].

A proper lesion characterization allows us to choose the best treatment according to patient status. In fact, although surgical resection is the only curative treatment in pancreatic ductal adenocarcinomas (PDAC), when the lesion is detected in an early phase, with regard to PNEN, surgical treatment should be in G1 and G2 NEN, although several authors suggested a watch-and-wait approach for asymptomatic non-functional lesions <2 cm. However, surgery is the front line treatment in young patients in cases of local invasiveness and in the presence of functioning lesions. In this scenario, the necessity of a proper lesion characterization is evident, although biopsy remains an invasive approach and problematic for small and deep lesions.

In the presence of unclear imaging results in the diagnosis of PNEN, CT radiomics can be helpful in determining the type of tumor. In fact, PNECs in the portal venous phase on CT with contrast media tend to show greater uniformity and less entropy than PDACs in the absence of significant differences in these parameters in the arterial phase, kurtosis, and asymmetry (heterogeneity parameters) [63]. The cut-off value of 0.34 has been suggested for uniformity in differentiating NEC from PDAC with a sensitivity and specificity of 79% and 65%, and a cut-off value of 1.89 for entropy with a sensitivity of 74% and a specificity of 70% [63]. Radiomics is also particularly useful in differentiating atypical PNENs that show hypovascularization in the CT arterial phase from PDACs [64]. In fact, the latter have a greater skewness (which is a measure of heterogeneity) than atypical PNETs with lower mean, median, fifth, tenth, and twenty-fifth percentiles on the entire tumor structure by contrast-enhanced CT (CECT) compared to those of atypical PNET. The authors suggest that this result could be linked to a greater cystic necrosis and degeneration. The fifth percentile alone or combined with asymmetry was capable in differentiating the two groups with high sensitivity: when considered alone, a sensitivity of 96% and a moderate specificity of 64% were reached, while in combination, a sensitivity of 90% and a specificity of 80% were reached [64]. In another study, it emerged that a model of seven radiomic characteristics can differentiate between atypical PDAC and PNET with greater sensitivity and specificity than the model based on clinical radiological parameters alone [65]. Neuroendocrine tumors can be differentiated from adenocarcinomas by histogram analysis performed in DWI of apparent diffusion coefficient (ADC) values on MRI (Figure 4). PDACs showed greater kurtosis (heterogeneity marker) and skewness in ADC histogram analysis on ADC₄₀₀ (b value 0–400 s/mm²) and ADC₈₀₀ (b value 0–800 s/mm²) than PNET tumors, while neuroendocrine tumors, as seen in the analysis of the CT texture, have a significantly lower entropy, regardless of the b value [66].

In addition to PDAC, radiomics seems capable of identifying other pathological entities that go into differential diagnosis with the PNETs, requiring different diagnostic-therapeutic pathways.

Among all machine-learning classifiers on CT-based radiomics analysis, Random Forest (RF) seems to be an excellent classification algorithm for differentiating PNETs from pancreatic cystadenomas with high accuracy, sensitivity, and specificity values (0.983, 0.980, and 0.986, respectively) [67]. Radiomic features of MRI may discriminate PNENs from pancreatic pseudopapillary neoplasms (SPTs), which often have similar imaging features, especially when small [68]. Furthermore, Shi et al. developed a radiomics model that incorporated the sex and age of patients and the radiomics signature of the tumor with excellent discrimination performance for diagnosing SPTs and PNETs with an area under curve (AUC) of 0.97 and 0.86 in a primary and validation cohort, compared to median diffusivity (MD) and median kurtosis (MK) diagnostic performance (area under curve of 0.71 and 0.65, respectively) [69].

More generally, radiomic models outperformed the clinical radiological model in discrimination against both PDACs and other lesions, reaching even higher values when a holistic model incorporated clinical and radiomics features was developed [68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88].

Tumor Grading

Most GEP-NEN radiomics studies focus on the evaluation of tumor grading in pNENs [89].

PNENs, according to the 2010 WHO classification system, were divided into three grades for their mitotic count and Ki-67 index: G1, low-grade (mitotic count < 2 per 10 high power fields (HPFs) and/or a Ki67 index of <3%); G2, intermediate-grade (mitotic count 2–20 per 10 HPF and/or a Ki67 index of 3–20%); and G3, high-grade (mitotic count > 20 per 10 HPF and/or a Ki67 index of >20%). These are also called poorly differentiated neuroendocrine carcinomas (pNECs). The 2017 WHO update reports that a subset of pNET G3 was identified with survival times shorter than those for pNET G1 or G2 but longer than typically described for pNEC (5-year survival rates varied from 60 to 100% for G1/2 and 16 to 29% for G3 pNEN) [90]. Treatment decisions for patients with pNETs are usually guided after staging the disease assessment. According to the European Society for Medical Oncology (ESMO) guidelines [91], treatment planning for pNEN G1 and G2 may involve only surgical resection with a sparing approach, while for more undifferentiated lesions, systemic therapy may be required [91]. So, the ability to distinguish the tumor grade of NETs preoperatively could have a huge impact on patient management and several radiomics features demonstrated to correlate with this prognostic factor [92].

Entropy, uniformity, and kurtosis extracted by CT imaging were significantly different between G1/2 tumors and G3 tumors [93,94,95,96]. Moreover, at high sigmoid levels, kurtosis was significantly different between G1 and G2 tumors. In addition, entropy was the most sensitive parameter to differentiate between G1/2 PNET and G3 PNEC, obtaining a sensitivity of 91% and a specificity of 85%. In another study, the entropy resulted in the only feature capable of predicting tumor grade (accuracy = 65%) [97]. Among the different variables, similar to CT radiomics, entropy resulted in the best predictor of tumor grade and aggressiveness, even on MRI images [93,98,99,100,101,102,103,104,105]. In addition, an increase of asymmetry and kurtosis values was shown with an increase of tumor grade [93]. The entropy of ADC was significantly higher in G2 and G3 tumors than in well-differentiated G1 (cut-off of 6.6) tumors, with a sensitivity and specificity of 83.3% and 61.1%, respectively, and an AUC of 0.75 [98].

Studies with a significant Radiomics Quality Score (RQS) and application of several machine-learning algorithms demonstrated a higher-order feature correlation with a prediction of pNET tumor grade [106,107]. On CT imaging, a higher value of “GLSZM Small Area High Gray-Level Emphasis” on the portal venous phase was recorded in G1 (0.80  ±  0.90) compared to G2/3 tumors (− 0.47  ±  0.86; p  < 0 .05) [107], while several features, such as GLCM_Homogeneity and GLCM_Entropy_log10, were selected from RF and discriminative of G3 tumors [106].

Some authors developed prediction models for PNET grading [108,109,110,111] built from combinations of multiple features extracted from radiological images, with encouraging results. In one study [108], researchers formulated a radiomic score based on four high-order selected features (compactness1.shape.original, LongRunHighGrayLevelEmphasis.GLRLM.original, SmallAreaHighGrayLevelEmphasis.GLSZM.squareroot, and skewness.firstorder.wavelet.LLH) extracted from the portal venous phase, with an overall high accuracy in tumor grading even in small pNET lesions (AUC = 0.86 for all PNETs; AUC = 0.81 for PNETs ≤ 2 cm). In contrast, Choi et al. [109] constructed a logistic regression model with the radiologic features of 66 pNET patients (45 patients with G1 tumors and 21 patients with G2/3 tumors) and the textural features of the arterial and venous phases leading to CT being capable of identifying the most undifferentiated lesions; this model is better suited for tumor grade determination (AUC = 0.77) than simple CT (AUC = 0.68). In line with these results, Canellas et al. [97] established a logistic regression model that was able to predict grading with an accuracy of 79.3%, considering entropy in the portal phase with CT imaging features of pNET (tumor diameter, vascular invasion, pancreatic duct dilation, lymphatic metastasis). However, higher performance is recorded from holistic models based on radiomics, clinical, and radiological features. Gu et al. [110], in a multicenter study, proposed a clinical/radiomic model, including tumor margin and the fusion radiomic signature from the arterial and portal venous phase CT images, demonstrating an optimal performance in both training and validation cohorts (accuracy value of 89.4% and 76.5%, respectively). Similarly, Liang et al. [111] combined clinical radiological data and CT radiomic features in a model, which resulted in being able to differentiate tumor grading in both training and validation sets with high accuracy (training set: AUC = 0.907; validation set: AUC = 0.891).

3. Results

An appealing target of radiomics in an oncological setting is the possibility to successfully predict a prognosis through the computational analysis of multiple preoperative parameters [39,40,41,42]. The ultimate purpose is to obtain individualized clinical decision-making and to increase patients’ survival rates [112,113,114,115].

Radiomics consists of the extraction of multiple features from medical images using mathematical algorithms. These features may have the potential to uncover tumor patterns and characteristics that cannot be appreciated with the naked eye. Volumes of raw data produced by CT, MRI, PET/CT, or even PET/MRI are used to find different pixel/voxel characteristics that can be useful for predicting prognoses and therapeutic responses for various types of cancer, thus providing valuable insights for therapy and prognosis [39,40,41,42].

The presence of lymph node metastasis plays a substantial prognostic role in pancreatic and small bowel (SB) NETs. Changes in the cytokine milieu, the expansion of immunosuppressive cell lineages, and an increase in lymphangiogenesis all contribute to the seeding, survival, and subsequent expansion of tumor cells within the lymph node. The surgical resection of tumor-draining lymph nodes has come under scrutiny, especially in more superficial tumors, where an extended lymph node dissection is often associated with substantial morbidity for patients [6,7]. In SB-NETs, however, regional mesenteric lymphadenectomy has been associated with improved survival. Nonetheless, the extent of lymphadenectomy has come under question. Patients with too few resected lymph nodes can be understaged, whereas the resection of a greater number of lymph nodes may enable the detection of clinically silent lymph node metastases. It is known that the limits of imaging in evaluating the real involvement of nodes is a diameter <15 mm; however, radiomics is in the early stages with regard to this feature [18].

So, several authors have focused on the performance of radiomics in predicting the recurrence and long-term clinical outcome in GEP-NETs.

3.1. Prediction of Survival

Chen et al. [116] developed and validated a computed tomography (CT)-based method to predict the efficacy of sunitinib in patients with pancreatic NET. Tumor shrinkage > 10% at the first follow-up after sunitinib treatment was significantly associated with longer progression-free survival (PFS; p < 0.001) and was used as the primary treatment outcome. The authors then developed a radiomic signature that showed significantly higher AUC in the training (0.915) and validation (0.770) sets than the ratio of CT values. The proposed radiomics model accurately predicted tumor shrinkage and progression-free survival in pancreatic NET patients treated with sunitinib, and might help select patients suitable for sunitinib treatment.

Regarding the 5-year survival status, a machine-learning algorithm and DL demonstrated excellent results (AUC ranging from 0.87 to 0.90), as highlighted in two studies [116,117] on a large population sample with external validation cohorts. These reports [116,117] are developed on data retrospectively extracted from the large database of the Surveillance, Epidemiology, and End Results (SEER) registry, using the AJCC seventh staging system as the gold standard for prognosis assessment, giving these conclusions a certain reliability. Testing survival rates among GEP NET, Werner et al. [118] observed that entropy was independently associated with PFS and overall survival (OS), while skewness resulted in being independently associated with OS.

3.2. Prediction of Tumor Aggressiveness

Martini et al. [119] further outlined an inverse correlation between entropy and OS and a direct correlation between skewness and OS in a radiomics model based on CT imaging.

In contrast, in 68GaSSA PET/CT radiomics analysis, a higher entropy value resulted in a better association with PFS and OS, [120] in patients treated with PRRT. The entropy and kurtosis of the ADC were found to be higher in aggressive tumors, especially those with vascular invasion, lymph node disease, and liver metastases [98] (Figure 5). Although several parameters were significant for several aggressivity markers, kurtosis was found to be the best parameter in the identification of vascular invasion, showing an AUC of 0.763 using a cut-off value equal to 4.13, and in the identification of distant metastases, showing an AUC of 0.820 using a cut-off value of 3.64 [98].

3.3. Prediction of Recurrences

In terms of disease recurrence appraisal, clinical data may significantly impact when added to radiomic features, as shown by one study [121] on 225 GP-NETs. Considering radiomics scores and clinical pathological factors (age, Ki-67 index, tumor pathological type, tumor primary site, and TNM stage), An et al. [121] were able to properly assess the non-recurrent group and the recurrent group more accurately than the individual models, which analyzed independently (combined model with an AUC of 0.824; clinical data model with an AUC of 0.786, and radiomics model with an AUC of 0.712).

Different methods for recurrence assessment in GP-NENs [92] focused on KI67, which is an important prognostic indicator. In recent years, several deep-learner-based approaches have emerged to evaluate Ki-67 indices [122]. Vesterinen et al. [123] built a deep-learning-based Ki-67 proliferation index PI algorithm (KAI) that objectively calculates Ki-67 PI, while Govind et al. [124] were able to generate a comprehensive view of the tumor via a Ki-67 index-based heatmap. Radiomics models [123,124] showed an excellent correspondence with pathologists’ assessment on postoperative specimens or preoperative biopsy, which are hampered by being invasive and not automatically representative of the whole lesion. The possibility to capture the complex intratumoral heterogeneity is an important tool in intercepting the tumor’s overall biological behavior [125,126,127,128,129]. From a clinical perspective, these open up the possibility to select patients and guide the right treatment choice, allowing a more personalized approach.

4. Prospects and Limits

Radiomics has already been demonstrated to improve tumor diagnosis, grading and staging, evaluation of responses to therapy, and prognosis prediction [130,131,132,133,134,135]. Machine-learning and deep-learning algorithms have supported fast processes of quantitative data analysis, further increasing the application of radiomics in oncological settings [136,137,138].

Additionally, radiogenomics can be used to create imaging biomarkers that can identify the genomics of a disease without the use of a biopsy, and various techniques are used to find statistically significant correlations between MRI, CT, and PET-imaging features and disease genomics [136,137,138].

The possibility to obtain biological information without an invasive approach is particularly attractive [139,140,141,142,143,144,145]. GEP-NETs are rare tumors with varied biological behavior [1,2,3]. Intercepting radiomics features useful in the management of these clusters of patients is of great scientific interest.

As noted in this review, most studies concerned the prediction of tumor grade [93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111]. The use of radiomics in the prediction of histological and genetic characteristics in different types of PNEN-evaluating images, both quantitatively and objectively, with texture analysis is growing in this field of research. From each voxel of a selected tissue region, hundreds of first-order quantitative data, such as entropy, and/or second-order characteristics, such as gray level, are extracted and analyzed with different post-processing methods and software. First-order features linked to increased heterogeneity, such as entropy and kurtosis [93,98,99,100,101,102,103,104], resulted in more undifferentiated tumors [93,94,95,96,97,98]. These data could be tested on GEP-NETs other than pancreatic type. Entropy was also frequently reported as a prognosis and recurrence predictor [45,47], and might be worth further exploring. First-order statistics were also the best predictors for distinguishing PNETs from PDACs [63,64]. Ultimately, the best performance was achieved when radiomics features were combined with clinical features [69,110,111,121]. All this evidence still needs to be reproduced in a larger cohort of patients and in a prospective manner to ensure reliability.

The standardization and retrospective nature of these studies are critical issues [145,146,147,148,149,150,151,152,153,154,155,156,157]. Of note, most of the available reports are focused on pancreatic NETs, and almost completely lacking any assessment of prognosis and long-term outcomes [18]. This is probably related to the rarity of these lesions and the difficulty of recruiting patients, which is reflected in research with small samples and external validation sets that are not always available. A separate validation dataset is an important element to avoid overfitting, which is another critical issue in radiomics surveys [146,147,148,149,150,151,152,153,154,155,156,157,158]. Additionally, current radiomics methods are limited to the use of single image sets for radiomics feature extraction and may not be able to capture the true features of the underlying tissues in the high-dimensional multiparameter-imaging space.

Consequently, in most cases, reports do not meet quality standards, and the predictive power of radiomics may be overestimated.

5. Conclusions

Although the evidence is promising, radiomics in GEP-NETs is still in its early stages. It has been preliminarily demonstrated that several radiomic features are correlated with some prognostic factors in GEP-NETs, but the combined clinical and radiomic models appear to achieve greater accuracy in lesion assessment. Radiomic models outperformed clinic-radiological model, both in discrimination against PDACs and other lesions, reaching even higher values when a holistic model incorporated clinical and radiomics features is developed. Regarding the 5-year survival status, Machine learning algorithm and DL demonstrated excellent results so as in tumor aggressiveness and recurrence. However, in most cases, the reports do not meet quality standards, and the predictive power of radiomics may be overestimated.

In the future, noninvasive predictive models may assist clinicians to optimize tailored therapeutic strategies for each patient with GEP-NETs in daily practice.