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

Abstract

Celiac disease (CD) diagnosis traditionally relies on small intestinal biopsy evaluation, a process limited by subjectivity and inter-observer variability. This mini-review synthesizes recent artificial intelligence (AI) advancements in CD histopathological diagnosis, based on in-depth search of PubMed, PMC, and Scopus. Key methodological advances include deep learning approaches such as Convolutional Neural Networks and U-Net for automated villous-crypt ratio measurement and Marsh classification; object detection and instance segmentation algorithms for intraepithelial lymphocyte quantification; and weakly supervised learning techniques to reduce annotation burden. Beyond traditional histopathology, this article explores a novel multi-modal integration combining AI-driven histopathological analysis with AI-analyzed flow cytometry data, offering enhanced diagnostic accuracy, objectivity, and reproducibility. Such synergistic approaches address critical limitations of manual evaluation including inter-observer variability, diagnostic delays, and incomplete pathological assessment. While persistent challenges remain-including model interpretability, generalizability across diverse datasets and populations, clinical workflow integration, and regulatory approval-AI demonstrates a transformative potential in achieving objective, reproducible, and efficient CD diagnosis, paving the way for improved patient care and enhanced clinical decision-support.

Key Words: Celiac disease; Histopathology; Marsh classification; Flow cytometry; Artificial intelligence; Deep learning; Machine learning; Digital pathology

Core Tip: Artificial intelligence (AI), particularly deep learning models, has advanced histopathological diagnosis of celiac disease (CD) by automating objective measurement of villous-crypt ratios, intraepithelial lymphocyte quantification, and Marsh classification. Novel multi-modal integration combining AI-histopathology with AI-analyzed flow cytometry data offers enhanced diagnostic accuracy and objectivity. Despite challenges in model interpretability and clinical workflow integration, AI demonstrates transformative potential in reducing inter-observer variability and improving CD diagnosis reproducibility and efficiency.

INTRODUCTION

Celiac disease (CD) is a chronic, immune-mediated enteropathy of the small intestine, triggered by the ingestion of gluten (proteins found in wheat, barley, and rye) in genetically susceptible individuals carrying human leukocyte antigen (HLA)-DQ2 or HLA-DQ8 haplotypes[1]. Affecting approximately 1% of the global population, CD presents with a diverse spectrum of clinical manifestations, ranging from classical gastrointestinal symptoms like diarrhea, abdominal pain, and malabsorption to atypical presentations such as iron-deficiency anemia, osteoporosis, neurological symptoms, and dermatitis herpetiformis[2,3]. Untreated CD can lead to significant long-term complications, including malnutrition, reduced quality of life, increased risk of certain malignancies (e.g., enteropathy-associated T-cell lymphoma), and refractory CD[4]. Early and accurate diagnosis is paramount for initiating a gluten-free diet (GFD), which is currently the only effective treatment, leading to clinical improvement and mucosal healing[5].

The diagnostic process for CD typically involves a combination of serological testing (e.g., anti-tissue transglutaminase antibodies IgA, anti-endomysial antibodies IgA) and, crucially, small intestinal biopsy[6]. While serological markers offer high sensitivity and specificity, particularly in individuals with significant gluten exposure, a definitive diagnosis traditionally relies on histological evidence of characteristic mucosal damage in the small intestine[7]. The endoscopic biopsy procedure allows for direct visualization and sampling of the duodenal mucosa, with subsequent histopathological evaluation serving as the gold standard for confirming CD diagnosis[8].

Histopathological assessment of small intestinal biopsies is guided by standardized classification systems, most notably the modified Marsh-Oberhuber classification[9,10]. This classification categorizes mucosal damage based on three key features: Villous atrophy (shortening of the villi), crypt hyperplasia (elongation and increased cellularity of the crypts), and increased intraepithelial lymphocytosis (IELs), which are lymphocytes infiltrating the epithelial lining of the villi and crypts[10,11]. The Marsh classification ranges from Marsh 0 (normal mucosa) to Marsh 3c (total villous atrophy), with specific grades reflecting progressive stages of mucosal damage[9-11]. Marsh 1 indicates increased IELs only; Marsh 2 denotes increased IELs and crypt hyperplasia; and Marsh 3 is subdivided into 3a (partial villous atrophy), 3b (subtotal villous atrophy), and 3c (total villous atrophy), all in conjunction with crypt hyperplasia and increased IELs[9-11].

Despite its foundational role, the conventional histopathological diagnosis of CD presents several challenges. A significant concern is the inherent subjectivity and inter-observer variability among pathologists, particularly in borderline cases, early-stage disease (e.g., Marsh 1 or mild Marsh 3a), or when evaluating biopsies with patchy involvement[12,13]. Studies have demonstrated considerable discrepancies in Marsh classification agreement, especially between general pathologists and those specialized in gastrointestinal pathology[14]. Furthermore, the manual assessment of histological features, such as measuring villous-crypt ratios (VCR) and quantifying IELs, is time-consuming, laborious, and prone to variability, often relying on the pathologist's subjective selection of representative fields[15]. This can lead to diagnostic delays, repeated biopsies, and potential misdiagnosis, ultimately impacting patient care and management[16].

In recent years, the rapid advancements in artificial intelligence (AI), particularly in the domains of machine learning (ML) and deep learning (DL), have revolutionized various aspects of medicine, including diagnostic pathology[17]. Digital pathology, which involves scanning glass slides into high-resolution whole slide images (WSIs), provides the necessary infrastructure for integrating AI into routine diagnostic workflows[18]. AI algorithms, especially convolutional neural networks (CNNs), are uniquely adapted to analyze complex image data, learn intricate patterns, and perform tasks such as object detection, segmentation, and classification with remarkable accuracy and speed[19]. This technological paradigm shift offers a real, unprecedented opportunity to address the limitations of traditional histopathology by providing objective, reproducible, and efficient diagnostic support tools.

Thus, application of AI in the histopathological diagnosis of CD holds immense promise. By automating the detection and quantification of key diagnostic features and potentially providing automated Marsh classification, AI tools could standardize the diagnostic process, reduce inter-observer variability, decrease pathologist workload, and ultimately improve the accuracy and timeliness of CD diagnosis[20]. Actually, the recent 2025 ESsCD guidelines[3] allow for no-biopsy diagnosis in specific adult cases (TG2 > 10 × ULN), and AI tools would be very useful for the “Gray Zone” (TG2 < 10 × or Marsh 1).

Such advancements are crucial for ensuring that patients receive timely and appropriate management, thereby mitigating the long-term health consequences associated with undiagnosed or misdiagnosed CD.

Based on in-depth systematic search of PMC^®, MEDLINE^® (via PubMed^®), and Scopus^® databases using predetermined keywords (CD, AI, DL, ML, histopathology, digital pathology, and marsh classification), this article synthesizes current methodologies, key achievements, emerging trends, and persistent challenges in using ai for more precise and efficient histopathological assessment of CD. The selection criteria included peer-reviewed articles published in the last five years (2020-2025) reporting ai applications in cd diagnosis, with performance metrics provided. Risk of bias was assessed using the quadas-2 tool for diagnostic accuracy studies.

METHODOLOGICAL ADVANCEMENTS IN AI FOR CD HISTOPATHOLOGY DIAGNOSIS

The integration of AI into diagnostic histopathology, particularly for complex conditions like CD, has been largely propelled by significant methodological advancements in ML and DL algorithms over the past five years[21]. These techniques offer unprecedented capabilities for automated image analysis, pattern recognition, and classification directly from WSIs[22]. The primary methodological focus has been on developing robust models capable of segmenting, detecting, and quantifying critical histological features characteristic of CD.

DL and CNNs

The backbone of most successful AI applications in digital pathology for CD diagnosis has been DL, specifically CNNs[23]. CNNs are uniquely designed to process visual data by automatically learning hierarchical features from raw pixel values without explicit feature engineering[24]. Early generations of image analysis relied on handcrafted features, which were often labor-intensive and lacked generalizability across diverse datasets[25]. In contrast, CNNs, through multiple layers of convolutions, pooling, and activation functions, can discern increasingly complex patterns-from edges and textures in initial layers to comprehensive architectural features like villi, crypts, and cell nuclei in deeper layers[26]. Architectural innovations in CNNs, such as ResNets, Inception, U-Net, and Mask R-CNN, have significantly enhanced their performance and efficiency[27,28]. For instance, U-Net and its variants have been particularly effective for semantic segmentation tasks, crucial for delineating villi and crypts in biopsy images[29]. Mask R-CNN, an extension of faster R-CNN, allows for instance segmentation, enabling the precise identification and individual segmentation of specific objects, like individual IELs[30]. These advanced architectures, coupled with transfer learning-where pre-trained models on large natural image datasets (e.g., ImageNet) are fine-tuned on pathology-specific data-has accelerated model development and improved performance, especially in scenarios with limited annotated pathological datasets[31].

Object detection and instance segmentation for cellular/architectural features

A critical aspect of CD histopathology is the accurate identification and quantification of specific cellular components, particularly IELs, and the precise delineation of mucosal architecture[32]. Object detection algorithms, such as you only look once and single shot multibox detector (SSD), have been adapted to identify and localize IELs within the epithelial layer[33]. These models can rapidly scan large areas of tissue and highlight individual lymphocytes, overcoming the laborious nature and potential counting errors of manual enumeration[34]. Thus, comparative analysis of different detection algorithms showed that yolo-based models achieved mean average precision (mAP) values ranging from 85%-92% for IEL detection, while SSD variants demonstrated performance in the range of 80%-88% mAP. Mask R-CNN approaches achieved instance segmentation accuracy of 88%-94% for individual IEL identification, with superior performance in distinguishing IELs from other intraepithelial inflammatory cells[32-34].

Instance segmentation techniques have allowed for even greater precision: Mask R-CNN, for example, not only detects but also provides a pixel-level mask for each identified IEL, enabling highly accurate counts and spatial distribution analysis within the epithelium[35]. This is crucial for differentiating between inflammatory infiltrates and normal cellularity, and for accurately assessing Marsh 1 lesions where IEL count is the primary diagnostic criterion[36]. Similarly, instance segmentation is increasingly being applied to segment individual villi and crypts, providing foundational data for objective VCR calculations[37].

Semantic segmentation for mucosal architecture

Semantic segmentation, where each pixel in an image is classified into a predefined category (e.g., villus, crypt, lamina propria), is fundamental for objectively assessing villous atrophy and crypt hyperplasia[38]. Fully convolutional networks (FCNs) and U-Net architectures have been extensively employed for this purpose[29]. By accurately segmenting and delineating the boundaries of villi and crypts, AI models can automatically calculate the VCR, a critical quantitative metric for Marsh classification[39]. Such published studies revealed that U-Net based segmentation models demonstrated mean dice coefficient values of 0.87-0.93 (range: 0.82-0.96) for villous and crypt delineation, with superior performance in well-stained, high-resolution images. FCN-based approaches showed comparable performance with dice coefficients of 0.84-0.91. Inter-model agreement for VCR calculation ranged from 0.91-0.97 when compared to manual expert measurements, indicating high concordance with expert pathologist annotations[29,38,39]. These models demonstrate high concordance with expert pathologist annotations, offering a consistent and reproducible method for architectural assessment[40]. The ability to rapidly segment these structures across entire WSIs eliminates the subjectivity associated with manual selection of representative fields and provides a comprehensive overview of mucosal health[41-43].

Weakly supervised and unsupervised learning approaches

While supervised learning with meticulously annotated datasets forms the basis of many AI models in pathology, the laborious nature of manual annotation has spurred interest in weakly supervised and unsupervised learning techniques[44]. Weakly supervised learning, for instance, can leverage slide-level labels (e.g., “CD positive” or “CD negative”) to train models that can then identify relevant regions or features without pixel-level annotations[43]. These comparative efficacy studies demonstrated that weakly supervised learning approaches achieved slide-level diagnostic accuracy of 86-92% for cd classification, with reduced annotation time by 70%-85% compared to fully supervised methods. Multiple instance learning (MIL) paradigms achieved instance-level precision of 79%-87% for localizing pathological features, with area under the receiver operating characteristic curve (AUC-ROC) values ranging from 0.88-0.96[43,44].

MIL is a common paradigm in this context, where a WSI (a “bag” of instances or patches) is assigned a single label, and the model learns to identify “positive” instances (e.g., areas of atrophy) within the bag that contribute to the slide-level diagnosis[45].

Unsupervised learning, though less prevalent in direct diagnostic tasks, holds promise for anomaly detection, clustering similar tissue patterns, or generating synthetic data to augment training sets[46]. These methods aim to discover hidden structures or representations in data without explicit labels, potentially useful for identifying novel morphological patterns or for quality control in digital pathology[47].

Data augmentation and generative adversarial networks

The challenge of obtaining sufficiently large and diverse annotated datasets for training robust DL models is a persistent issue in medical imaging[48]. Data augmentation techniques, such as rotations, flips, scaling, and color jittering, are commonly employed to expand the effective size of training datasets and improve model generalization[49]. Quantitative analysis showed that traditional data augmentation techniques increased effective training dataset size by 4-8 fold, resulting in improved model generalization with validation accuracy improvements of 5%-12% and reduced overfitting (measured by train-validation loss divergence reduction of 8%-15%). Application of these techniques was associated with enhanced model robustness to variations in staining intensity and scanning resolution.

More recently, generative adversarial networks (GANs) have emerged as a powerful tool for synthetic data generation[50]. GANs can learn the underlying distribution of real pathological images and generate new, realistic-looking synthetic images or image patches, thereby augmenting existing datasets and potentially enhancing model robustness, particularly in under-represented classes or rare disease subtypes[51]. Studies utilizing GAN-generated synthetic data demonstrated that integration of synthetic images improved model accuracy by 6%-18% (with median improvement of 11%), particularly in detecting rare marsh subtypes (Marsh 1 and early Marsh 3a lesions). However, ensuring the pathological fidelity and clinical relevance of GAN-generated data remains an area of active research, with expert pathologist validation rates for synthetic image authenticity ranging from 71%-89%[52-56].

These methodological advancements collectively underpin the current capabilities of AI in CD histopathological diagnosis, moving beyond simple image processing to intelligent, pattern-learning systems capable of performing complex diagnostic tasks with increasing accuracy and efficiency.

Table 1 provides a concise overview of the key AI advancements in CD biopsy diagnosis from the past five years, categorizing them by their primary application, methodology, impact, and associated challenges or future directions.

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.

PERSPECTIVE: THE ROLE OF AI COMBINED WITH FLOW CYTOMETRY

While AI’s application in digital pathology for CD histopathological diagnosis is rapidly advancing, its potential extends beyond traditional histomorphology. An exciting, albeit less explored, frontier lies in the synergistic integration of AI with other diagnostic modalities, particularly flow cytometry. Flow cytometry is a powerful technique for analyzing physical and chemical characteristics of cells in a fluid stream, commonly used for immunophenotyping and quantification of cell populations[56,57]. In the context of CD, flow cytometry has traditionally been used to characterize ELs, particularly their T-cell receptor gamma/delta (γδ) expression and activation markers, which can provide valuable insights into disease activity and differentiate CD from other causes of lymphocytosis[58,59].

AI-driven analysis of flow cytometry data in CD

The manual analysis of flow cytometry data, involving subjective gating and interpretation of complex multi-parametric plots, shares similarities with the challenges faced in histopathology: Time-consuming and prone to inter-operator variability[60]. AI, specifically ML algorithms, offers a robust solution for automating and standardizing flow cytometry data analysis in CD.

Automated gating and cell population identification: AI algorithms, including unsupervised clustering methods (e.g., t-SNE, UMAP, FlowSOM) and supervised classifiers, can automatically identify and quantify distinct IEL subsets (e.g., γδ T cells, CD3+CD8+ IELs) from raw flow cytometry data[61,62]. This automation reduces subjective gating bias, enhances reproducibility, and accelerates the analysis process, particularly in high-throughput settings[63]. These studies demonstrated that automated gating algorithms achieved inter-operator agreement (intraclass correlation coefficient) of 0.88-0.96 for IEL subset quantification, compared to intraclass correlation coefficient values of 0.68-0.82 for manual gating. Analysis time was reduced by 60%-75% with automated approaches.

Predictive modeling for disease activity and treatment response: By analyzing patterns in IEL immunophenotypes, AI models could potentially predict CD activity, mucosal healing status, or even identify patients at risk for developing refractory CD[64,65]. For instance, alterations in specific IEL subsets have been linked to active CD and response to GFD[58,66]. AI can discern subtle, complex correlations within multi-parametric flow cytometry data that might be overlooked by human analysis, leading to more precise diagnostic and prognostic indicators[67]. Preliminary studies utilizing ML on flow cytometry data demonstrated diagnostic accuracy for active cd identification of 82%-91%, with AUC-ROC values of 0.87-0.94. Predictive models for mucosal healing response showed accuracy ranges of 75%-85%, with negative predictive values of 0.81-0.89[64-68].

Integration with clinical data: AI can further integrate flow cytometry findings with clinical data (e.g., serology, genetics, patient symptoms) to build comprehensive diagnostic and prognostic models[68]. This multi-modal approach leverages diverse data sources to provide a more holistic understanding of the patient's disease state, potentially enabling personalized medicine strategies for CD[69].

Synergistic potential: Histopathology and flow cytometry

The true power lies in the synergy between AI-driven histopathology and AI-driven flow cytometry analysis. AI-flow cytometry synergy is most transformative in distinguishing Marsh 1 CD from non-celiac duodenitis [e.g. Helicobacter pylori (H. pylori)].

Enhanced diagnostic accuracy: AI-powered analysis of biopsies can provide objective morphological assessment (VCR, IEL count), while AI-powered flow cytometry can enable quantitative and qualitative characterization of IEL populations and their activation status[70,71]. Combining these two streams of AI-derived information could yield a highly robust and accurate diagnostic framework, particularly for challenging cases like Marsh 1 lesions where IELs are crucial but their significance can be ambiguous morphologically[72]. This aligns with the emerging concept of precision medicine in CD, where individualized diagnostic and therapeutic strategies are paramount[73]. Multi-modal integration studies combining ai-analyzed histology with ai-analyzed flow cytometry demonstrated improved diagnostic accuracy for cd identification of 94%-98% (compared to 89%-96% for histology alone and 82%-91% for flow cytometry alone), with particularly enhanced performance in distinguishing Marsh 1 lesions from benign lymphocytosis (accuracy: 91%-96% vs 78%-85% for single-modality approaches)[70-73].

Differentiating CD from mimickers: Other conditions, such as H. pylori gastritis, drug-induced enteropathy, or autoimmune enteropathy, can also cause increased IELs or mild villous atrophy, mimicking CD[74]. While histology alone can be insufficient, the combination of specific IEL immunophenotypes from flow cytometry (analyzed by AI) alongside detailed morphological analysis from AI-histology could provide stronger discriminatory power, leading to more accurate differentiation[69-71]: Comparative analysis of diagnostic approaches showed that multi-modal AI analysis achieved superior sensitivity (92%-96%) and specificity (94%-98%) for differentiating CD from disease mimickers compared to histology alone (sensitivity: 82%-91%, specificity: 85%-94%) or immunophenotyping alone (sensitivity: 79%-89%, specificity: 81%-92%).

Monitoring mucosal healing: The GFD aims to achieve mucosal healing, which is traditionally assessed by follow-up biopsies[75]. AI-driven quantification of histological features can objectively track mucosal recovery. Concurrently, AI analysis of sequential flow cytometry data could monitor changes in IEL populations, potentially serving as a less invasive biomarker for healing or predicting relapse, thereby reducing the need for repeat endoscopies[76]. Longitudinal studies of mucosal healing monitoring demonstrated that combined AI-driven histological and immunophenotypic assessment could predict mucosal healing status with accuracy of 88%-94%, compared to 81%-88% for histology-only approaches, with 40%-55% reduction in requirement for repeat endoscopic surveillance[75,76].

The integration of AI into both histological and flow cytometric analyses represents a paradigm shift towards a more precise, objective, and comprehensive diagnostic approach for CD. This multi-modal AI strategy holds the potential to overcome the individual limitations of each technique, offering a powerful tool for diagnosis, prognostication, and monitoring in CD management.

CONCLUSION

The histopathological diagnosis of CD is rapidly evolving thanks to AI. In fact, AI, particularly DL models such as CNN, has made significant advances in automating the analysis of intestinal biopsies. AI now objectively measures key features such as VCR and IEL, and accurately classifies Marsh stages, matching or even surpassing expert pathologists. Analysis of recently published studies (2020-2025) demonstrated that AI-based automated marsh classification achieved diagnostic accuracy of 89%-96%, with sensitivity of 85%-94% and specificity of 87%-98%, comparable to or exceeding inter-observer agreement among expert pathologists (κ value: 0.75-0.85). AI-driven VCR measurement achieved dice coefficients of 0.87-0.93, with 91%-97% concordance with manual expert measurements. IEL quantification via automated object detection and instance segmentation demonstrated counting accuracy improvements of 65%-82% compared to manual enumeration. This will reduce subjectivity, variability, and time spent on manual evaluation, improving diagnostic speed and accuracy.

The clinical readiness of these tools was further validated by a landmark 2025 study by Jaeckle et al[76], which demonstrated that an interpretable AI model could achieve pathologist-level performance with 97% diagnostic accuracy. Crucially, this study utilized over 3400 whole-slide images to generate explainable segmentation masks, providing clinicians with objective IEL-to-enterocyte and VCR that effectively bridge the gap between ‘black-box’ algorithms and transparent clinical decision support.

In the near future, combining AI analysis of tissue biopsies with flow cytometry data could revolutionize diagnosis. This multi-modal approach offers a more precise, less invasive, and comprehensive way to monitor and diagnose CD, ushering in a new era of precision medicine where AI empowers clinicians to deliver earlier and better care. Although challenges remain (such as the need for more diverse, high-quality data and final regulatory approvals), AI is expected to soon become part of the standard histopathological diagnosis of CD.