Artificial intelligence in autoimmune pancreatitis: Differentiating immunoglobulin G4-related pancreatitis from pancreatic ductal adenocarcinoma

Table 1 Potential applications of artificial intelligence in differentiating autoimmune pancreatitis from pancreatic ductal adenocarcinoma

AI application	Data source	Potential role	Clinical value	Key limitations
CT radiomics	Pancreatic-protocol CT; multiphasic contrast-enhanced CT	Quantitative assessment of lesion texture, shape, enhancement, and heterogeneity	May help distinguish focal AIP from PDAC when conventional CT findings overlap	Retrospective studies, small AIP cohorts, scanner variability, segmentation dependence[6-9,18]
MRI/DWI-based AI	MRI, MRCP, diffusion-weighted imaging, ADC maps	Analysis of ductal anatomy, soft-tissue contrast, diffusion restriction, and parenchymal features	Useful when CT is equivocal or when additional non-invasive characterization is required	Variation in MRI protocols, b-values, motion artifacts, and reproducibility[11,25]
PET/CT radiomics	18F-FDG PET/CT	Quantitative assessment of metabolic uptake patterns and spatial heterogeneity	May provide additional information in complex or equivocal cases where inflammation and malignancy both show FDG uptake	FDG uptake is non-specific; PET/CT is not routinely required in all suspected AIP or PDAC cases[10,19]
AI-assisted EUS image analysis	EUS images or videos	Detection and characterization of solid pancreatic lesions; support for real-time image interpretation	May reduce operator dependence and support selection of lesions requiring tissue acquisition	Limited AIP-specific validation; image quality and operator technique influence performance[26-28]
AI-supported tissue acquisition	EUS-guided biopsy data; lesion imaging features	Identification of targets for repeat biopsy or areas more suspicious for malignancy	May help reduce sampling error in indeterminate pancreatic masses	Requires prospective validation; cannot replace adequate histology[16,26,27]
Digital pathology AI	Biopsy or resection slides; IgG4 immunostaining	Detection of malignant glands, inflammatory infiltrates, fibrosis patterns, and IgG4-positive plasma-cell distribution	May support distinction between lymphoplasmacytic sclerosing pancreatitis and malignancy	AIP-specific pathology datasets are limited; small biopsies may lack architecture[1,4,16,29,30]
Multimodal AI models	CT/MRI radiomics, EUS features, histology, IgG4, CA19-9, bilirubin, other-organ involvement	Integrated risk prediction for AIP vs PDAC	Closest to real-world multidisciplinary decision-making; might support steroid-trial selection or repeat-biopsy decisions	Complex model development; missing data, interpretability, external validation, and cancer-safety concerns[6,8,31,32]
Explainable AI and decision support	Model outputs, heatmaps, feature importance, risk scores	Transparent presentation of AI predictions to clinicians	Might improve trust, allow error checking, and support multidisciplinary discussion	Explainability methods may be imperfect and require clinical validation[33,34]

AI: Artificial intelligence; ADC: Apparent diffusion coefficient; AIP: Autoimmune pancreatitis; CA19-9: Carbohydrate antigen 19-9; CT: Computed tomography; DWI: Diffusion-weighted imaging; EUS: Endoscopic ultrasound; FDG: Fluorodeoxyglucose; IgG4: Immunoglobulin G4; MRCP: Magnetic resonance cholangiopancreatography; MRI: Magnetic resonance imaging; PDAC: Pancreatic ductal adenocarcinoma; PET: Positron emission tomography.

Table 2 Reported diagnostic performance of published artificial intelligence studies for differentiating autoimmune pancreatitis or mass-forming pancreatitis from pancreatic ductal adenocarcinoma

Ref.	Scope	Cohort (cases vs PDAC; total)	Modality	AI method	Validation	Performance	Key limitation
Park et al[6]	AIP vs PDAC	89 AIP/93 PDAC; n = 182	CT	Radiomics + random forest	Hold-out test	AUC 0.975; Sn 89.7%; Sp 100%	Single-center; no external validation
E et al[7]	Focal AIP vs PDAC	45/51; n = 96	Multiphasic CT	Radiomics	No independent validation	Sn 93.3%; Sp 96.1%; AUC NR	Small; single-center
Li et al[8]	Focal AIP vs PDAC	42/55; n = 97 (matched)	CT	Radiomics (LASSO score)	Whole matched cohort	AUC 0.97; Sn 95.2%; Sp 92.7%	Optimism risk; no independent validation
Lu et al[9]	Focal AIP vs PDAC	32/64; n = 96	CT	Radiomics nomogram	5-fold CV + hold-out test	Test AUC 0.83; Sn/Sp NR	Small; no external validation
Liu et al[10]	AIP vs PDAC	48/64; n = 112	Dual-time 18F-FDG PET/CT	Radiomics + SVM	Nested cross-validation	AUC 0.967; Sn 85.3%; Sp 96.0%	CV only; needs dual-time PET/CT
Shiraishi et al[11]	Focal AIP vs PDAC	28/77; n = 105	MRI (DWI/ADC)	Radiomics + SVM	Cross-validation	AUC 0.962; Sn/Sp NR	Small AIP subgroup; single-center
Anai et al[23]	Focal AIP vs PDAC	20/30; n = 50	CT texture	Texture ML (SVM) + reader study	3-fold CV	SVM AUC 0.920; Sn 100%; Sp 75%; readers 0.83-> 0.91	Very small; single-center
Ziegelmayer et al[20]	AIP vs PDAC	44/42; n = 86	CT (portal venous)	Deep CNN features	Shuffle-split CV + 20% hold-out	AUC 0.90; Sn 0.89; Sp 0.83	Small; high variance; no external validation
Zhang et al[19]	AIP vs PDAC	n =111	18F-FDG PET/CT	Radiomics + SVM	Nested cross-validation	AUC 0.93; Sn 0.86; Sp 0.84	CV only; cohort breakdown unclear
Ye et al[22]	Mass-forming pancreatitis vs PDAC	78/120; n = 198	CT	Radiomics + clinical	Hold-out test	Test AUC 0.906; Sn/Sp NR	Not AIP-specific; single-center
Qu et al[24]	Mass-forming pancreatitis vs PDAC	54/201; n = 255	CT	Radiomics + clinical	Internal validation	Validation AUC 0.880; Sn/Sp NR	Not AIP-specific; internal validation only

Performance refers to the best-performing model on the validation or test set where available; several studies reported the area under the curve only. Mass-forming pancreatitis studies are included for context and are not specific to autoimmune pancreatitis. AI: Artificial intelligence; AIP: Autoimmune pancreatitis; PDAC: Pancreatic ductal adenocarcinoma; CT: Computed tomography; MRI: Magnetic resonance imaging; DWI: Diffusion-weighted imaging; ADC: Apparent diffusion coefficient; PET: Positron emission tomography; FDG: Fluorodeoxyglucose; SVM: Support-vector machine; CNN: Convolutional neural network; CV: Cross-validation; AUC: Area under the curve; Sn: Sensitivity; Sp: Specificity; NR: Not reported.