Role of artificial intelligence in the characterization of indeterminate pancreatic head mass and its usefulness in preoperative diagnosis

Abstract

Artificial intelligence (AI) has been used in various fields of day-to-day life and its role in medicine is immense. Understanding of oncology has been improved with the introduction of AI which helps in diagnosis, treatment planning, management, prognosis, and follow-up. It also helps to identify high-risk groups who can be subjected to timely screening for early detection of malignant conditions. It is more important in pancreatic cancer as it is one of the major causes of cancer-related deaths worldwide and there are no specific early features (clinical and radiological) for diagnosis. With improvement in imaging modalities (computed tomography, magnetic resonance imaging, endoscopic ultrasound), most often clinicians were being challenged with lesions that were difficult to diagnose with human competence. AI has been used in various other branches of medicine to differentiate such indeterminate lesions including the thyroid gland, breast, lungs, liver, adrenal gland, kidney, etc. In the case of pancreatic cancer, the role of AI has been explored and is still ongoing. This review article will focus on how AI can be used to diagnose pancreatic cancer early or differentiate it from benign pancreatic lesions, therefore, management can be planned at an earlier stage.

Key Words: Artificial intelligence; Indeterminate pancreatic lesion; Imaging; Biomarkers; Diagnosis

Core Tip: Surgical management of a pancreatic head lesion usually requires pancreaticoduodenectomy, which is associated with significant morbidity and mortality. For a benign lesion it is unacceptable. The investigation modalities i.e. computed tomography, magnetic resonance imaging, endoscopic ultrasound, positron emission tomography, and biochemical markers are available today to distinguish benign from malignant lesions and have their limitations (human judgmental errors). The application of artificial intelligence (AI) algorithms can minimize human errors and improve the sensitivity and specificity of diagnostic yield. The AI can help with great precision in differentiating benign from malignant lesions, affecting the management strategy and minimizing post-operative complications.

INTRODUCTION

The concept of a machine that can think like a human being was proposed by Mr. Alan Turing in the year 1950 in his book entitled “Computing Machinery and Intelligence” and later, the term “artificial intelligence (AI)” was coined by John McCarthy[1,2]. The applicability of AI ranges from simple tasks to more complex tasks mimicking a human brain. There are six major sub-fields of AI: machine learning (ML), neural network, deep learning (DL), natural language processing (NLP), cognitive computing, and computer vision. ML can learn from data, recognize typical patterns, and make decisions with little or no human interference. A neural network is the field of AI that is inspired by the human brain, where a set of algorithms is used to derive a correlation. Most of the AI models in the medical field use ML and neural networks. NLP is a method where textual data has been used to search, analyze, and comprehend complex information. Computer vision understands visual inputs (radiological or pathological images, surgical videos) and derives desired information. There are many modifications of conventional sub-fields of AI which have been in use. The twentieth century has seen that AI has become an essential part of day-to-day life, including health tracking devices[3], automobiles[4], banking and finances (robo-traders)[5], surveillance, social media, entertainment, education, space exploration, and disaster management, etc[6,7].

AI has been used in various fields of medicine including online appointments and hospital check-ins, medical records digitalization, follow-up, drug dosage reminders, adverse effect warnings, etc. Moreover, its application in the field of oncology is paramount. AI can be useful in cancer detection, screening, diagnosis, classification, prognostication, new drug discovery, etc[8-11]. It has played its role in differentiating various indeterminate lesions in the thyroid gland[12,13], breast[14], lungs[15,16], liver[17], adrenal[18,19], kidneys[20], and indeterminate biliary strictures[21] (Table 1). Various authors have studied the role of AI algorithms to identify pancreatic lesions from imaging modalities computed tomography (CT), magnetic resonance imaging (MRI), endoscopic ultrasonography (EUS), positron emission tomography (PET) scan, etc and thus can differentiate malignant indeterminate pancreatic lesions (IPLs) from benign ones for better management at an early stage.

Table 1 Studies on differentiation of indeterminate lesions using artificial intelligence.

No.	Ref.	Number of patients	Organ of interest	Sub-type of AI	Outcome
1	Ippolito et al[12], 2004	453	Thyroid nodule (benign vs malignant)	ANN	Refinement of risk stratification of FNAB and clinical data
2	Daniels et al[13], 2020	121	Indeterminant thyroid nodule	ML	ML and ultrasonography can identify genetically high risk lesions
3	Becker et al[14], 2018	632	Breast lesion (benign vs malignant)	Generic DLS	Aids diagnosing cancer on breast ultrasound images with an accuracy comparable to radiologists
4	Scott et al[15], 2019	125	Lung GGO (benign vs malignant)	ANN	Improve diagnostic ability using CT scan, PET, and clinical data
5	Guo et al[16], 2022	20	Indeterminant small lung lesions	DNN	DNN based method may detect small lesions < 10 mm at an effective radiation dose < 0.1 mSv.
6	Yasaka et al[17], 2018	460	Liver mass (HCC vs others)	CNN	High diagnostic performance in differentiation of liver masses using dynamic CT
7	Moawad et al[18], 2021	40	Adrenal incidentaloma (benign vs malignant)	ML	Machine learning and CT texture analysis can differentiate between benign and malignant indeterminate adrenal tumors
8	Stanzione et al[19], 2021	55	Indeterminant solid adrenal lesions	ML	MRI handcrafted radiomics and ML can be used to different adrenal incidentalomas
9	Massa'a et al[20], 2022	160	Indeterminant solid renal mass (benign vs malignant)	ML	MRI-based radiomics and ML can be useful in differentiation
10	Saraiva et al[21], 2022	85	Indeterminant biliary strictures	CNN	CNN can accurately differentiate benign strictures from malignant ones

AI: Artificial intelligence; ANN: Artificial neural network; CNN: Convolutional neural network; CT: Computed tomography; DNN: Deep neural network; DLS: Deep learning software; FNAB: Fine needle aspiration biopsy; GGO: Ground glass opacities; HCC: Hepatocellular carcinoma; ML: Machine learning; MRI: Magnetic resonance imaging.

IPLs are those detected by imaging techniques performed for non-specific abdominal complaints or detected incidentally, otherwise known as pancreatic incidentaloma. With the increase in imaging modalities, the detection of such IPLs has increased[22]. These incidentalomas are mostly detected in other organs, i.e. the thyroid gland, pituitary gland, kidney, lungs, adrenal gland, etc. Though, the incidence of indeterminate lesions is less in the pancreas, however, most of them are malignant compared to other sites[23]. Identification of such lesions creates confusion in clinicians and anxiety among the patients. Moreover, early diagnosis of malignancy can provide reasonably early management and better overall outcomes. Therefore, it is necessary to diagnose such lesions for better patient management.

The overall prevalence of such lesions was reported to be 0.01%–0.6% in 2009, which may be less compared to its true incidence[24]. A review of a series of pancreatic resections shows an asymptomatic neoplastic lesion to be 6%-23% (24% to 50% of them are malignant, and 24% to 47% are considered potentially malignant or pre-malignant)[25,26]. A recently published Leopard-2 trial comparing laparoscopic and open pancreaticoduodenectomy has shown the incidence of benign or pre-malignant lesions to be 12%[27]. Frequently, cystic lesions of the pancreas are detected on MRI and their incidence is up to 20%[28] and recent series shows the incidence to be 49% in the general population[29]. The majority of cystic lesions are benign, however, approximately, 3% are malignant or potentially malignant[30].

The etiology of such lesions is diverse, benign adenoma to adenocarcinoma, borderline malignant tumors, mesenchymal tumors, neuroendocrine tumors, cysts, congenital changes, metastatic lesions, inflammatory masses etc[23]. These lesions may be broadly divided into benign, pre-malignant, or malignant lesions[24]. Figure 1 shows different pathologies of IPLs[31].

Figure 1 Pathology of different indeterminate pancreatic lesions.

There is a considerable overlap of imaging features of different benign and malignant pancreatic lesions. Cystic degeneration of solid tumors may masquerade as cystic lesions. Various modalities (ultrasonography, contrast-enhanced CT, MRI, EUS, PET, cytopathology, histopathology, and tumor markers) have been used to differentiate the possible etiology, however, there are limitations of each modality intrinsic to the investigation itself or on the operator. Recently, AI has been used to distinguish various indeterminate lesions in the breast, lungs, adrenal gland, kidney, etc. Thus, the use of AI in association with conventional imaging or diagnostic modalities can improve their overall diagnostic yield and therefore, more precise diagnosis and patient care.

This paper reviews the current status of AI in the differentiation of various IPLs and its future implications.

METHODS AND LITERATURE SEARCH

All the relevant articles were searched from PubMed and Google Scholar using the keywords, i.e. “artificial intelligence” AND “pancreatic lesions” OR “cystic lesions”, OR “CT”, OR “MRI”, OR “EUS”, OR “PET” OR “pathology”, OR “biomarkers” between 2005 and 2023, and only full articles were studied. Articles discussing the differentiation of different types of pancreatic lesions were included and screened by all authors. Abstracts and conference presentations were excluded. Studies discussing the differentiation of any pancreatic lesion (benign vs. malignant) were included in relevant sections for discussion. The study flow chart is shown in Figure 2.

Figure 2 Study flow chart.

Role of clinical parameters and AI on the identification of IPLs

Pancreatic cancer is one of the leading causes of cancer-related death worldwide, thus early diagnosis is crucial for better management. Often, patients are asymptomatic to start with, so presentation is delayed leading to advanced disease at diagnosis. This delay in diagnosis can be minimized by the identification of high-risk groups and the introduction of targeted screening of high-risk populations. Any lesion identified in these patient groups can be subjected to further evaluation using an AI augmented imaging system (CT, MRI, PET, EUS), which will be discussed later. The proposed schema of patient evaluation and management is presented in Figure 3.

Figure 3 Schematic presentation of diagnosis of indeterminate pancreatic lesion using artificial intelligence. AI: Artificial intelligence; CT: Computed tomography; EUS: Endoscopic ultrasonography; MRI: Magnetic resonance imaging; PET: Positron emission tomography; SOL: Space occupying lesion.

Several clinical parameters can be used to predict the future incidence of pancreatic cancer including, symptoms, hereditary factors (BRCA1, BRCA2, PALB2, Hereditary pancreatitis, and Peutz-Jeghers Syndrome), pre-existing clinical conditions (new-onset diabetes mellitus), lifestyle (smoking, alcohol, obesity, nutrient-poor diet), and demographic factors. Elevation of CA 19-9, CEA, and recently developed CEMIP (cell migration-inducing hyaluronan binding protein) can be considered as an early indicator of pancreatic cancer[32-34]. None of these parameters can confirm pancreatic cancer, however, a combined assessment can suggest a possible pancreatic cancer leading to screening of high-risk populations. In a retrospective study from Kaiser Permanente Southern California, an algorithm for risk stratification for pancreatic cancer was generated using imaging (CT/magnetic resonance) and clinical factors[35]. In this study, imaging features used were pancreatic duct dilatation as a predictor of malignancy and other features such as atrophy, calcification, pancreatic cyst, and irregular pancreatic duct. Multi-state prediction model showed a discriminatory index (c-index: 0.825–0.833) between normal individuals and individuals with pancreatic cancer. A study at the Biomedical Imaging Research Institute of Cedars Sinai Medical Center, Los Angeles used ML and CT-based radiomic features as an indicator of pancreatic ductal adenocarcinoma (PDAC)[36]. The scans were obtained in non-pancreatic cancer patients for different purposes, who later developed pancreatic cancer after 6 mo to 3 years. The AI model had an accuracy of 86% in the prediction of PDAC. As CT scans were performed frequently for different purposes, such AI models can identify patients having potential risk for future pancreatic malignancy.

Muhammad et al[37], Placido et al[38], and Chen et al[39] used demographic and clinical parameters with artificial neural networks (ANNs) algorithms to predict pancreatic cancer. In the validation arm, the area under the curve (AUC) was 0.85, and the sensitivity and specificity of diagnosis were 80.7%. Malhotra et al[40] used ML principles to identify symptoms to predict pancreatic cancer. Their algorithm could detect 41.3% of patients with pancreatic cancer < 60 years of age, 20 mo earlier than diagnosis (AUC: 0.66), and 43.2% of patients with pancreatic cancer > 60 years of age, 17 mo earlier than diagnosis (AUC: 0.61). Appelbaum et al[41] used neural network algorithms to identify high-risk groups 1 year in advance. Thus, these AI techniques not only help to detect pancreatic cancer but also, earlier than conventional imaging.

Role of AI on CT scan imaging on detection of pancreatic lesions

If a mass lesion is detected in the pancreas, the possibility of neoplasm is kept as a differential diagnosis. The most common (85%–95%) among the lesions is pancreatic ductal adenocarcinoma (PDAC) and it has a poor prognosis[42,43]. Ill-defined hypovascular mass is the characteristic of PDAC in contrast-enhanced imaging[44]. Atypical imaging of a solid mass may harbor a malignancy, however, its mimic, an inflammatory mass, can have a better prognosis than PDAC, and management of both these conditions is different.

Among all the imaging modalities, CT is most commonly favored for the investigation of a pancreatic lesion, as it is widely available, quick to acquire, has a high spatial resolution, assesses relationship to vascular structures, and determines surgical planning. Recent advances in CT imaging in the form of multiplanar reformatted images, and three-dimensional (3D) techniques have improved sensitivity by up to 96% in tumor identification[45,46]. However, small tumors or tumors with atypical features may not be visible on CT scans or subtle changes may not be appreciable to the human eye and prone to errors. These limitations of conventional CT imaging can be overcome by the use of AI algorithms.

Differentiation of PDAC

Among all malignancies, PDAC has the worst overall survival[47]. It is because patients present late at an advanced stage due to late detection of asymptomatic subtle pancreatic lesions on imaging[40]. Zhu et al[48] and Liu et al[49] have used DL to detect pancreatic cancer and in the study by Liu et al[49], malignancy could be detected in just 3 s with an AUC of 0.96. Chu et al[50] could diagnose PDAC with an AUC of 99.9% using ML algorithms.

Differentiation of cystic lesions

With the increase in the frequency of cross-sectional imaging, the detection of cystic lesions of the pancreas has increased and it is aptly called “technopathies”. Management of these cystic lesions requires classification of the type of lesion and the risk of malignancy which is sub-optimal with present imaging modalities[51,52]. AI has been used to differentiate the types of cystic lesions into, intraductal papillary mucinous neoplasm (IPMN), mucinous cystic neoplasm (MCN), serous cystic neoplasia (SCN), solid pseudopapillary neoplasia, etc[53,54]. Dmitriev et al[53] used the convolutional neural network (CNN) model (contrast-enhanced CT and clinical data) to differentiate the types of cystic lesions with an accuracy of 84% which is better than radiologists which has an accuracy of less than 70%[53,55]. However, Li et al[54] used only CT images and AI (DL) to differentiate the cystic lesions with an accuracy of 73% compared to radiologists in their study which had an accuracy of only 48%. Differentiation of SCN from other cystic lesions is important as they have a rare chance of being malignant, thus, Wei et al[56] used an ML-based algorithm to distinguish SCN from others based on CT images. Yang et al[57] and Chen et al[58] have used AI algorithms to distinguish SCN from MCN. Chakraborty et al[59] and Polk et al[60] used the RF model to differentiate low-grade IPMN from high-grade IPMN which has management implications. Table 2 summarizes studies on the uses of AI along with CT images in the differentiation of pancreatic lesions.

Table 2 Studies on differentiation of indeterminate lesions using artificial intelligence algorithms on computed tomography images.

No.	Ref.	Number of patients	Primary objective	Sub-type of AI used	Outcome
1	Qureshi et al[36], 2022	108	Identification of PDAC	ML	Accuracy: 86%
2	Ebrahimian et al[121], 2022	103	Differentiation of benign vs malignant pancreatic lesions	RF	AUC: 0.94
3	Chakraborty et al[59], 2018	103	High risk vs low risk IPMN	RF, SVM	AUC: 0.81
4	Polk et al[60], 2020	29	High risk vs low risk IPMN	LR	AUC: 0.90
5	Ikeda et al[122], 1997	71	PDAC vs pancreatitis	NN	AUC: 0.916
6	Chen et al[58], 2021	100	SCN vs MCN	LASSO and RFE_Linear SVC	AUC: 0.932
7	Yang et al[57], 2019	53	SCN vs MCN	LASSO	AUC: 0.66
8	Yang et al[123], 2022	63	SCN vs MCN	MMRF-ResNet	AUC: 0.98
9	Ren et al[124], 2020	112	PDAC vs pancreatic adenosquamous carcinoma	RF	AUC: 0.98
10	Xie et al[125], 2021	226	MCN vs ASCN	RF	AUC: 0.734
11	Ziegelmayer et al[126], 2020	86	AIP vs PDAC	CNN, ML	AUC: 0.90
12	Li et al[62], 2022	97	Focal-type AIP vs PDAC	LASSO regression	AUC: 0.97
13	Gao et al[127], 2021	170	MCN vs SCN	mRMR + LASSO	AUC: 0.91
14	Dmitriev et al[53], 2017	134	Classification of pancreatic cyst	RF, CNN	Accuracy: 83.6%
15	Li et al[54], 2019	206	Classification of pancreatic cysts	DNN (Dense-Net)	Accuracy: 72.8%
16	Wei et al[56], 2019	260	SCN vs other cystic neoplasms	ML	AUC: 0.767

AI: Artificial intelligence; AIP: Autoimmune pancreatitis; ASCN: Atypical serous cystic neoplasm; AUC: Area under the curve; CNN: Convolutional neural network; DNN; Deep neural network; IPMN: Intraductal papillary mucinous neoplasm; LASSO: Least absolute shrinkage and selection operator; LR: Logistic regression; MCN: Mucinous cystic neoplasm; ML: Machine learning; PDAC: Pancreatic ductal adenocarcinoma; RFE: Recursive feature elimination; RF: Random forest; SCN: Serous cystic neoplasm; SVM: Support vector machine; NN: Neural network; mRMR: Minimum redundancy maximum relevance; SVC: Support vector classifier; MMRF: Multi-channel-multiclassifier-random forest.

Role of AI on MRI on the detection of pancreatic lesions

MRI is favored over CT scan due to superior soft tissue delineation and it also helps to detect small lesions, assessment of the vascular relationship, and relationship to the pancreatic duct, lymph node, or distant metastasis[43,61]. Detection of iso-attenuating pancreatic lesions on CT scan is challenging which is observed in approximately 10% of patients. In these situations, indirect evidence of malignancy is used for diagnosis, i.e. convex pancreatic contour, double duct sign, vascular involvement, mass effect, etc[42]. However, MRI can be helpful to diagnose such lesions. Recently, the use of AI algorithms has improved the diagnostic ability of MRI. Li et al[62] and Chen et al[63] used AI algorithms for the identification of PDAC on different phases of MRI (Table 3).

Table 3 Studies on differentiation of indeterminate lesions using artificial intelligence algorithms on magnetic resonance images.

No.	Ref.	Number of patients	Primary objective	Sub-type of AI used	Outcome
1	Li et al[62], 2022	267	PDAC detection	UDA + meta learning + GCN	DSC (62.08%, T1), (61.35%, T2), (61.88%, DWI), (60.43%, AP)
2	Chen et al[63], 2022	73	PDAC detection	Spiral-ResUNet	DSC: 65.60%, Jaccard index: 49.64%
3	Liang Y et al[128], 2020	56	PDAC detection	CNN	DSC: 71%
5	Cui et al[129], 2021	202	Grading-BD IPMN	LASSO	AUC: 0.903
6	Corral et al[67], 2019	139	Classification of IPMN	CNN	AUC: 0.783
7	Cheng et al[68], 2022	60	Malignant IPMN	LR, SVM	MRI + SVM: AUC: 0.940, CT + SVM: AUC: 0.864
8	Hussein et al[130], 2019	171	Classification of IPMN	SVM, RF, 3D, CNN	Accuracy 84.22%

AI: Artificial intelligence; AP: Arterial phase; AUC: Area under the curve; CT: Computed tomography; CNN: Convoluted neural network; DSC: Dice similarity coefficient; DWI: Diffusion weighted image; GCN: Graph convolutional network; IPMN: Intraductal papillary mucinous neoplasm; LASSO: Least absolute shrinkage and selection operator; LR: Logistic regression; MRI: Magnetic resonance and imaging; PDAC: Pancreatic ductal adenocarcinoma; RF: Random forest; SVM: Support vector machine; UDA: Unsupervised data augmentation.

Management of cystic lesions depends upon the precise characterization, which indicates its clinical behavior[64]. However, overlapping imaging features make differentiation challenging[64]. The role of imaging is to differentiate benign from malignant cystic neoplasms. MRI uses T2 images to identify ductal communication and post-contrast images to characterize the lesion. It is limited in the detection of calcifications which is better appreciated on a CT image. MRI can differentiate benign from malignant lesions with an accuracy of 73% to 81% compared to a CT scan which has an accuracy of 75% to 78%[52,65,66].

The use of AI has enabled MRI to detect high-grade dysplasia or malignancy in IPMN with a sensitivity and specificity of 75% and 78%, respectively[67]. Corral et al[67] used 3D CNN to classify IPMN into different types with an accuracy of 58%. Interestingly, Cheng et al[68] compared radiomics features of CT and MRI using AL algorithms [LASSO, LR, support vector machine (SVM)] and found out that, the MRI-based model(AUC: 0.940) had better diagnostic ability than the CT based model(AUC: 0.864). Studies on the use of AI with MRI to detect the type of cystic or solid pancreatic lesions are presented in Table 3.

Role of AI on EUS in the detection of pancreatic lesions

EUS uses a high-frequency transducer at the tip of an endoscope. It helps to obtain high-resolution images of the pancreas through the esophagus, stomach, or duodenum. Various modalities of EUS including contrast-enhanced EUS, EUS-guided fine needle aspiration (FNA), and EUS elastography have been used for the evaluation of pancreatic cancer, detection of small lesions, differentiation of solid from cystic tumors, and assessment of resectability[69]. Most importantly, it helps to obtain tissue for cytopathology or histopathology[70,71]. The main drawback is operator dependency, which may reduce the diagnostic yield[72,73]. AI algorithms have been used in association with EUS to detect pancreatic cancers and to differentiate from other lesions (Table 4). Mass-forming chronic pancreatitis may masquerade as pancreatic malignancy, EUS based AI algorithms can be used to distinguish pancreatic cancer from chronic pancreatitis.

Table 4 Studies on differentiation of indeterminate lesions using artificial intelligence algorithms on endoscopic ultrasonography images.

No.	Ref.	Number of patients	Primary outcome	Sub type of AI used	Outcome
1	Zhu et al[78], 2013	262	PDAC vs CP	SVM	Accuracy: 94.2%
2	Zhu et al[131], 2015	100	AIP vs CP	SVM	Accuracy: 89.3%
3	Zhang et al[74], 2010	216	Normal pancreas vs PDAC	SVM	Accuracy: 97.98%
4	Ozkan et al[76], 2016	332	Recognition of pancreatic cancer amongst various age group	ANN	Accuracy: Average: 87.5% (all ages), Min: 88.46% (40-60 yr), Max: 92% (< 40 yr)
5	Kuwahara et al[83], 2019	50	Benign vs malignant IPMN	CNN	Accuracy: 94%
6	Das et al[75], 2008	56	PDAC vs normal pancreas vs CP	ANN	AUC: 0.93
7	Săftoiu et al[80], 2008	68	Benign vs malignant pancreatic lesion	ANN	Accuracy: 89.7%
8	Tonozuka et al[132], 2021	139	PDAC vs CP	CNN	AUC: 0.94
9	Marya et al[133], 2021	583	PDAC vs benign causes of pancreatic SOL	CNN	AUC: 0.976
10.	Xu et al[134], 2013	Systemic Analysis of 6 studies	Benign vs malignant pancreatic lesion	-	AUC: 0.962

AI: Artificial intelligence; AIP: Autoimmune pancreatitis; ANN: Artificial neural network; CNN: Convoluted neural network; CP: Chronic pancreatitis; IPMN: Intraductal papillary mucinous neoplasm; PDAC: Pancreatic ductal adenocarcinoma; SOL: Space occupying lesion; SVM: Support vector machine.

Authors have used ML algorithms to differentiate normal pancreatic tissue from PDAC with more than 93% accuracy[74-76]. Two studies have used AI to distinguish chronic pancreatitis from PDAC on EUS images with an accuracy of more than 80%[77,78]. Săftoiu et al[79] demonstrated better diagnostic ability of contrast-enhanced EUS (94.6% and a specificity of 94.4%) compared to EUS-FNA (87.5% and 92.7%) in differentiating CP from PDAC using AI.

Recently, EUS elastography has been used to diagnose focal pancreatic lesions. Using ANN, it can differentiate benign from malignant lesions with an accuracy of 95%[80]. In another multicenter prospective study using ANN, they demonstrated that EUS elastography (sensitivity (87.6%) and specificity (82.9%)) had better diagnostic ability than two experienced endoscopists combined (sensitivity 80.0%, specificity 50.0%)[81]. Udriştoiu et al[82] used ML principles to distinguish focal pancreatitis from pancreatic mass (neuroendocrine tumor or PDAC) with an accuracy of 98.26%. Differentiation of benign IPMN from malignant IPMN has management implications, Kuwahara et al[83] studied to detect malignant IPMN using CNN (ResNet-50).

Role of AI on PET imaging on the detection of pancreatic lesions

PET is a functional imaging technique used for staging malignant lesions and is based on the physiological characteristics of tumor cells[84,85]. However, inflammation may mimic a malignant lesion due to high metabolic activity giving rise to false positive results, conversely, in patients with hyperglycemia, it can give a false negative result[86,87]. PET CT is also useful in the assessment of tumor response to therapy[43]. Li et al[88] used a hybrid feedback-SVM-random forest model to detect pancreatic cancer from a normal pancreas with an accuracy of 96.47%. Liu et al[89] studied the role of dual time PET/CT and SVM model to differentiate PDAC from AIP with an AUC of 0.96 similarly, Xing et al[90] showed a diagnostic performance of 0.93 of AUC.

Role of AI in pathological examination on detection of pancreatic lesions

Often, imaging cannot achieve an accurate diagnosis, requiring a tissue diagnosis-cytology or histology[91,92]. AI can be applied to hematoxylin-eosin-stained slides for the detection of pancreatic cancer[93]. Song et al[94] used AI algorithms to segment epithelial cell nuclei on slide images and extract morphological features and could differentiate SCN from MCN and grading of PDAC[95]. The CNN was used by Kriegsmann et al[96] to localize pancreatic intra-epithelial neoplasm or PDAC in a slide. Niazi et al[97] used DL to detect neuroendocrine tumors from normal tissues on Ki-67 stained biopsy images with a 97.8% sensitivity and 88.8% specificity. Momeni-Boroujeni et al[98] could differentiate benign from malignant pathology using a K-means clustering algorithm from FNA-based slides with an accuracy of 100%. Naito et al[99] used CNN in FNB-based slides to assess PDAC with an AUC of 0.984. Cyst fluid analysis is an essential part of the diagnosis of pancreatic cystic lesions. Kurita et al[100] used a neural network to differentiate benign from malignant cysts taking into consideration biomarkers in cyst fluid, cytology and clinical parameters.

Role of AI in biomarkers on detection of pancreatic lesions

Biomarkers act as an adjunct in diagnosis, prognosis, and screening for recurrence and they can be used for early diagnosis of tumors. However, in the case of pancreatic cancer, it lacks sensitivity and specificity for routine clinical practice[91,101,102]. Liquid biopsy is one of the recent developments in oncology, developed with the intent of detecting tumor cells from blood when biopsy cannot be obtained, or to assess tumor response to therapy (surgery or chemoradiotherapy) and assess genetic mutation. It includes three types of sampling of biological materials; which are circulating tumor cells (CTCs), circulating tumor DNA, and exosomes. CTCs have faced difficulties for years because of very low concentrations in many studies, which is 1–10 cells per 10-mL of blood (much lower than billions of hematopoietic cells) and short half-life (approximately from 1 to 2.4 h) in blood which poses difficulty in further study. AI can be used in the detection of disease from these biomarkers and various studies have explored AI algorithms for biomarkers for diagnosis[91,103]. Studies used exosomes[104-106], cell-free DNA[107], extracellular vesicles long RNA[108], proteins[109-112], and circulating microRNA[113] in association with AI for diagnosis of pancreatic cancer. Table 5 shows studies on the role of biomarkers and AI in the differentiation of pancreatic lesions.

Table 5 Studies on differentiation of indeterminate lesions using artificial intelligence algorithms on different biomarkers.

No.	Ref.	Number of samples	Type of biomarker used	Sub-type of AI used	Conclusion
1	Chen et al[104], 2019	28	Exosomes	LDA	Accuracy: 100%
2	Zheng et al[105], 2022	220	Exosomes	ANN	AUC: 0.86
3	Ko et al[106], 2017	28	Exosomes	LDA	Accuracy: 100%
4	Cristiano et al[107], 2019	34	Cell-free DNA	GBM	AUC: 0.86
5	Yu et al[108], 2020	501	extracellular vesicles long RNA	SVM	AUC: 0.96
6	Gao et al[109], 2012	199	Proteomes	SVM, KNN, ANN	AUC: 0.971
7	Yu et al[110], 2005	100	Proteomes	DT	Sensitivity: 88.9%, specificity: 74.1%
8	Qiao et al[112], 2022	136	Proteomes	CNN	Accuracy: 87.63%
9	Alizadeh et al[113], 2020	671	Circulating micro RNA	ANN	Accuracy: 0.86

AI: Artificial intelligence; ANN: Artificial neural network; CNN: Convoluted neural network; DT: Digital transformation; KNN: K-nearest neighbor; GBM: Gradient boosting machine; LDA: Linear discriminant analysis; SVM: Support vector machine.

This review has shown that AI can be used in routine investigation modalities (CT, MRI, EUS, PET, biomarkers) to improve diagnostic and differentiating potential; however, it is still in progress. In the beginning, studies have trained and validated AI algorithms, in the future it is a challenge to implement such studies at different geographical locations, ethnicity, genetic makeup, etc. The majority of studies have explored the potential to differentiate, chronic pancreatitis from pancreatic ductal adenocarcinoma, SCN from MCN, and high-risk vs. low-risk IPMN, however, there can be other differential diagnoses in a clinical scenario.

DISCUSSION

Surgery for malignant pancreatic head lesions was standardized by Whipple et al[114] which is acceptable worldwide. It includes a complex single-stage procedure of pancreaticoduodenectomy, which is associated with morbidity (25%) and mortality (0%-9.3%) even in high-volume centers[115-117]. Professor Whipple[118] reported a mortality of 29.2% in his series of patients who underwent pancreaticoduodenectomy. Though, recent series have reported reduced mortality following pancreaticoduodenectomy, morbidity of the procedure continues to be high. Recently, many modifications have been made to reduce morbidity, however, none of the measures appeared to be successful. Are et al[119] reported a historical perspective where 7 out of 37 pancreaticoduodenectomies performed by Prof Whipple AO turned out to be chronic pancreatitis (18.9%), where such a morbid procedure could have been avoided. Recent series have also supported these findings of incidence of benign pathology following pancreaticoduodenectomy in the range of 5%-10%[117,120]. Hence, there is an unmet need to differentiate benign pancreatic lesions from malignant ones. Multiple imaging modalities have been used to distinguish benign from malignant lesions, however, each investigation modality has its limitations which are compounded by human errors. The application of AI has minimized those errors and can make diagnoses earlier. Table 6 shows how AI increases the yield of different imaging modalities for predicting a malignant pancreatic head lesion. We have proposed an algorithm for the diagnosis of such entities. Whenever a patient presents to a clinician, history and clinical examination precede imaging. Hence, AI can be used to develop algorithms to predict malignancy[32-34]. In a patient with a high risk of pancreatic malignancy, a pancreatic indeterminate lesion should be investigated further with imaging or biopsy to rule out malignancy. Studies have reported the usefulness of biomarkers in the diagnosis of pancreatic cancer[107-110]. Hence, all non-invasive markers (clinical, biochemical) can be used to develop an algorithm that can predict pancreatic cancer before imaging has been performed and it can differentiate malignant pancreatic lesions. As shown in Table 6, AI has an added advantage over conventional imaging in differentiating pancreatic cancer from benign conditions. So, those high-risk patients marked on non-invasive pancreatic cancer detection models can be subjected to AI-enhanced imaging for better diagnosis. Further in line, to clarify the final tissue diagnosis, AI can help to detect subtle markers that can be ignored by human error. Therefore, AI can be used in every step of the diagnosis of an indeterminate pancreatic head mass, to detect malignant lesions early thus, availing proper oncological management.

Table 6 Studies demonstrating impact of artificial intelligence on increasing efficacy of diagnostic modalities.

No.	Ref.	Objective	Modality	Sensitivity	Specificity	Accuracy
1	Corral et al[67], 2019	Differentiate cystic SOL of pancreas	Fukuoka guideline	62%	77	77.5%
			Deep learning	75%	78%	78.3%
2	Kuwahara et al[83], 2019	Detection of malignant IPMN	Human pre-operative diagnosis (Clinical + lab + imaging)	95.7%	22.2%	56%
			Artificial intelligence	95.7%	92.66	94%
3	Gao et al[135], 2020	Ability to differentiate pancreatic disease	CE-MR	NA	NA	83.93%
			GAN	NA	NA	76.79%
4	Rigiroli et al[136], 2021	Detection of pancreatic cancer and SMA involvement	CT scan	NA	NA	71%
			Artificial intelligence	62%	77%	54%
5	Chen et al[137], 2023	Detection of pancreatic cancer	CT scan	89.9%	95.9%	AUC: 0.96
			CNN	90%	93%	NA
6	Tang et al[138], 2023	Pancreatic mass diagnosis	EUS FNA	81.6%	100%	87.9%
			CE EUS Master-guided FNA	90.9%	100%	93.8%

CE-MR: Contrast enhanced-magnetic resonance; CT: Computed tomography; CNN: Convoluted neural network; EUS: Endoscopic ultrasound; FNA: Fine needle aspiration; GAN: Generative adversarial network; IPMN: Intraductal papillary mucinous neoplasm; NA: Not available; SMA: Superior mesenteric artery; SOL: Space occupying lesion.

Pancreatic incidentalomas or indeterminate lesions are on the rise due to the plethora of cross-sectional imaging performed to diagnose non-specific abdominal complaints. Though plenty of studies have been made in the fields of breast cancer, lung cancer, hepatocellular carcinoma, renal cell carcinoma, and adrenal tumors, there is a dearth of literature discussing how to differentiate benign pancreatic lesions from benign ones. The current literature included studies comparing individual pancreatic lesions, i.e. serous cystadenoma vs. mucinous cystadenoma, autoimmune pancreatitis vs. pancreatic adenocarcinoma, low-grade vs. high-grade IPMN, etc. However, a comprehensive review discussing how to differentiate various malignant pancreatic lesions (both cystic and solid) from benign lesions with the help of AI is lacking. Hence, in this review, we have discussed how to differentiate different pancreatic lesions encountered in day-to-day clinical practice using different algorithms of AI. We have discussed individually about different diagnostic modalities and different types of pancreatic lesions. There are more studies available in the field of radiological investigations and fewer studies available for the histopathological diagnosis or intra-operative differentiation of malignant from benign lesions. As the understanding of the usefulness of AI is increasing, these limitations can be curtailed in the near future.

FUTURE PERSPECTIVES

There is a surge in the number of medical imaging for different indications leading to the identification of many indeterminate pancreatic lesions (IPLs), which help to diagnose a disease earlier or can lead to a plethora of other investigations, psychological stress, clinical dilemmas, etc. Human judgment is prone to errors as subtle differences in these small or atypical lesions are challenging to discern leading to inter-observer and intra-observer variations which can be minimized with the use of AI.

CONCLUSION

AI is an evolving technical advancement in the field of medicine and can play a significant role in differentiating IPLs into benign or malignant, by enhancing the diagnostic yield of conventional imaging (CT, MRI, PET), EUS, tissue diagnosis (cytopathology, histopathology), and biomarkers (liquid biopsy). An early and accurate diagnosis may lead to timely intervention, thereby improving the patient outcome. The current literature on this is still limited and sparse, therefore, more studies are required to reach a standard approach for the application of AI in IPLs.