Artificial intelligence in celiac disease pathology: From digital histomorphology to multi-modal integration

Table 1 Synthesis of literature data and key artificial intelligence advancements in celiac disease histopathological diagnosis

AI application	Primary AI methodology	Key impact on CD histopathological diagnosis	Performance metrics1	Ref.	Challenges/future directions
Automated VCR measurement	Semantic Segmentation (U-Net, FCN), CNNs	Objective, reproducible quantification of villous atrophy and crypt hyperplasia; reduces inter-observer variability and manual effort.	Dice coefficient: 0.87-0.93; VCR Agreement with experts: 0.91-0.97; mean absolute error in VCR: 0.08-0.15	[39,41]	Robustness to diverse staining/scanning protocols; fine-tuning for subtle partial atrophy; validation across multicenter datasets
IEL quantification	Object detection (YOLO, SSD), instance segmentation (mask R-CNN)	Accurate, rapid counting and density estimation of IELs; aids in Marsh 1 diagnosis and differentiation from other lymphocytosis causes.	YOLO mAP: 85%-92%; SSD mAP: 80%-88%; mask R-CNN instance accuracy: 88%-94%; counting error reduction: 65%-82% vs manual enumeration	[34,36]	Differentiation of IELs from other inflammatory cells; consistency across varying tissue preparations; clinical correlation with specific IEL thresholds and establishment of standardized reference ranges
Automated marsh classification	CNNs (ResNet, Inception), MIL	Enables automated or semi-automated grading of mucosal damage (Marsh 0-3c); high sensitivity/specificity comparable to expert pathologists	Overall diagnostic accuracy: 89%-96%; sensitivity: 85%-94%; specificity: 87%-98%; AUC-ROC: 0.92-0.98; agreement with expert consensus (κ value): 0.85-0.94	[23,45]	Interpretability of “black-box” models (XAI); generalizability across unseen datasets/ethnicities; handling of equivocal or patchy lesions; ethical considerations regarding AI in clinical decision-making
Early disease detection/subtlety enhancement	Attention mechanisms in CNNs, anomaly detection	Highlights subtle changes in architecture or IELs in early CD; acts as a “second reader” to reduce missed diagnoses	Sensitivity for Marsh 1 detection: 82%-91%; sensitivity for early Marsh 3a: 79%-88%; Reduced missed diagnosis rate: 35%-48% vs single-reader evaluation	[20,53]	Validation against long-term patient outcomes for early detection; integration into pathologist workflow for seamless alert generation; prospective clinical trials
Digital pathology integration and workflow	Cloud-based AI platforms, API integration	Streamlines workflow from WSI acquisition to AI analysis and reporting; enables scalability and remote access.	Processing time reduction: 65%-80% vs manual review; WSI analysis throughput: 20-40 slides/hour; user satisfaction scores: 7.8-8.9/10	[18,54]	Interoperability with existing systems; data privacy and security compliance (General Data Protection Regulation, Health Insurance Portability and Accountability Act); infrastructure requirements for WSI storage and processing; cost-effectiveness analyses
XAI	Grad-CAM, LIME, attention maps	Increases transparency of AI decisions; builds trust and aids pathologists in understanding model rationale (e.g., heatmaps)	Pathologist concordance with XAI explanations: 81%-89%; Trust scores before vs. after XAI implementation: 5.2 → 7.8/10; Time to interpret AI output: Reduced by 40%-55%	[55,56]	Developing clinically meaningful explanations; standardization of XAI outputs for diagnostic review; ensuring XAI insights are actionable for pathologists; balancing complexity and interpretability: “Pathologist-in-the-loop” validation
Data augmentation/synthetic data generation	GANs, traditional augmentation techniques	Expands training datasets, improves model robustness and generalization, especially for rare findings or limited data	Dataset expansion: 4-8 fold with traditional augmentation; Model accuracy improvement: 6%-18% with GAN-generated data (median 11%); expert validation of synthetic images: 71%-89%	[49,51]	Ensuring pathological fidelity of synthetic data; preventing generation of misleading artifacts; ethical implications of using synthetic data in diagnostic training; regulatory framework development
Weakly supervised learning	MIL	Reduces annotation burden by leveraging slide-level labels for training; useful for large datasets where fine-grained annotation is impractical	Slide-level diagnostic accuracy: 86%-92%; annotation time reduction: 70%-85%; instance-level precision: 79%-87%; AUC-ROC: 0.88-0.96	[43,45]	Precision in localizing specific pathological features from weak labels; risk of model focusing on non-diagnostic features; validation of learned attention patterns; threshold establishment for clinical deployment

¹Performance metrics represent aggregated data from systematic review of literature (2020-2025) and represent ranges or median values where applicable. Metrics include standard measures such as sensitivity, specificity, accuracy, dice coefficient (for segmentation), mean average precision (mean average precision for detection), and area under the receiver operating characteristic curve. All values are derived from published peer-reviewed studies in PubMed^®, PMC^®, and Scopus^® databases.

CD: Celiac disease; AI: Artificial intelligence; VCR: Villous-crypt ratios; CNNs: Convolutional neural networks; GANs: Generative adversarial networks; AUC-ROC: Area under the receiver operating characteristic curve; IEL: Intraepithelial lymphocyte; YOLO: You only look once; SSD: Single shot multibox detector; MIL: Multiple instance learning; XAI: Explainable artificial intelligence; WSI: Whole slide images.