Frontiers | Artificial intelligence-assisted type 2 inflammatory endotyping in CRSwNP: from Sinus CT and digital pathology to biologic decision support

Abstract

Chronic rhinosinusitis with nasal polyps (CRSwNP) is a heterogeneous rhinologic disease frequently driven by type 2 inflammation, yet routine clinical phenotyping remains insufficient for endotype-informed management. Conventional markers, including blood eosinophils, serum immunoglobulin E (IgE), fractional exhaled nitric oxide, tissue eosinophilia, nasal polyp score, and sinus computed tomography (CT) scores, provide useful but incomplete information. Artificial intelligence (AI) enables integration of sinus CT, radiomic features, digital pathology, biomarkers, comorbidities, and follow-up treatment outcomes. Emerging studies suggest that CT-based AI supports non-invasive identification of eosinophilic or type 2-high disease, whereas digital pathology links routine histology with inflammatory and molecular endotypes. However, published evidence remains dominated by retrospective single-modality studies or partial multi-source integration, and no model has yet jointly integrated sinus CT, digital whole-slide pathology, and biomarkers in CRSwNP. Additional limitations include heterogeneous endotype labels, insufficient external validation, and uncertain clinical utility for biologic decision-making. This Mini Review synthesizes existing single-modality artificial intelligence pipelines, contextualizes CRSwNP as a representative disease model for AI-driven type 2 inflammatory endotyping, and proposes a clinically oriented, multimodal interpretable scoring framework integrating sinus imaging and digital pathology to support precision type 2 stratification in CRS management.

1 Introduction: chronic rhinosinusitis with nasal polyps as a clinically accessible model of type 2 inflammation

Chronic rhinosinusitis with nasal polyps (CRSwNP) is one of the most clinically recognizable and biologically heterogeneous phenotypes of chronic rhinosinusitis. In many Western populations and an increasing proportion of Asian patients, it is characterized by type 2 inflammation with eosinophilic infiltration and activation of interleukin (IL)-4, IL-5, IL-13, immunoglobulin E (IgE) and epithelial alarmin signaling (, ). The introduction of biologics targeting interleukin-4 receptor α (IL-4Rα), IgE, IL-5/interleukin-5 receptor (IL-5R), and upstream epithelial signals has shifted management from symptomatic surgery to mechanism-directed therapy, making CRSwNP an optimal model for precision rhinology (, ).

CRSwNP is also well suited to artificial intelligence (AI)-assisted endotyping, as routine clinical workflows yield multi-dimensional datasets: nasal endoscopy, sinus computed tomography (CT), surgical histology, measurable biomarkers, airway comorbidities and long-term therapeutic follow-up. Conventional phenotypes and single biomarkers incompletely capture this heterogeneity. Multimodal AI integrating radiomics, deep learning and digital pathology can synthesize multi-source data to generate interpretable inflammatory endotype risk scores.

This Mini Review frames CRSwNP as a rhinologic model of type 2 inflammation and summarizes AI tools spanning sinus CT, digital pathology and multimodal biologic decision-making. A targeted literature search was performed in PubMed and the Web of Science Core Collection from January 1, 2020 to June 2026. Four thematic domains were covered: AI-assisted sinonasal CT/imaging analysis, digital pathology/pathomics, multimodal or multi-omics machine learning, and AI-assisted biologic decision support or treatment-response prediction. Search terms combined rhinosinusitis, chronic rhinosinusitis, CRS, CRSwNP, and nasal polyps with AI-related terms, including artificial intelligence, machine learning, deep learning, convolutional neural network, radiomics, pathomics, digital pathology, whole-slide imaging, multimodal learning, multi-omics, ChatGPT, large language model, dupilumab, mepolizumab, omalizumab, and biologics. Eligible studies were peer-reviewed CRS/CRSwNP articles applying AI-related methods to four main domains mentioned above. Studies without an AI component, unrelated to sinonasal disease, conference abstracts, unpeer-reviewed preprints, purely preclinical mechanistic studies, and generic AI studies without rhinologic relevance were excluded. To reduce omission bias, broad keyword combinations were used consistently across both databases, reference lists of relevant articles and reviews were manually checked, and priority was given to studies with larger cohorts, multicenter data, external validation, model comparison, interpretability, and translational relevance. As this was a narrative Mini Review, no PRISMA flow diagram, formal risk-of-bias assessment, or quantitative synthesis was performed.

2 The impact of CRSwNP on quality of life and the rationale for AI-assisted type 2 endotyping in rhinology

AI is clinically relevant to CRSwNP because the disease itself is intrinsically multidimensional. In the biologic era, rhinologists must infer inflammatory mechanisms, recurrence risk, biologic eligibility, likely response, and treatment trajectories from symptoms, endoscopy, sinus CT, biomarkers, histopathology, comorbid airway disease, and prior treatment response. Separate analysis of each dataset cannot reliably identify inflammatory endotypes or deliver personalized therapeutic guidance.

2.1 Clinical phenotypes do not reliably define inflammatory endotypes

Although CRSwNP is defined by bilateral nasal polyps and characteristic sinonasal symptoms, phenotype does not reliably identify endotype. Patients with matching clinical manifestations vary widely in tissue eosinophilia, local IgE synthesis, neutrophilic infiltration, tissue remodeling and concurrent lower airway disease. Early biomarker-based clustering studies demonstrated that similar phenotypes encompass distinct inflammatory endotypes, including type 1 (T1), type 2 (T2), type 3 (T3), and mixed immune patterns (). Single-cell and transcriptomic studies further demonstrate immune, epithelial, and stromal heterogeneity within nasal polyps (, ). This heterogeneity also extends to aspirin-exacerbated respiratory disease (AERD) and non-AERD CRSwNP, which exhibit distinct immune endotypes and gene-expression profiles, further supporting molecular stratification beyond conventional clinical phenotyping (). Clinical phenotypes enable preliminary diagnosis yet cannot support mechanism-driven patient stratification.

2.2 Conventional biomarkers provide complementary but incomplete endotype information

Conventional biomarkers provide complementary but fragmented information. Blood eosinophils, serum IgE, allergen-specific IgE, and fractional exhaled nitric oxide reflect systemic type 2 inflammation but do not consistently represent local sinonasal disease. Tissue eosinophilia better predicts local disease prognosis yet requires surgical specimens and varies by sampling and counting standards. Clinical and imaging features, including nasal polyp score, Lund-Mackay score, ethmoid-to-maxillary dominance, prior surgery, corticosteroid exposure, asthma, nonsteroidal anti-inflammatory drug-exacerbated respiratory disease (N-ERD), and allergic rhinitis quantify disease severity but cannot forecast biologic response (, ). Critically, no single biomarker fulfills three core clinical needs: non-invasive preoperative screening, standardized postoperative endotyping and personalized biologic efficacy prediction. Therefore, AI-assisted inflammatory endotyping acts as an integrated probabilistic tool to synthesize scattered clinical data without replacing rhinologist clinical judgement.

3 CT-based AI for non-invasive inflammatory endotyping

3.1 Sinus CT as a routinely available substrate for AI-assisted inflammatory endotyping

Sinus CT is standard preoperative imaging for CRSwNP to quantify overall disease burden (). Although the Lund-Mackay score remains clinically useful (), it cannot capture high-dimensional information in sinus CT, including spatial distribution, density, texture, osteitis, and ethmoid-to-maxillary patterns (, ). AI can transform these features into quantitative biomarkers (). Emerging evidence identifies CT radiomic signatures correlated with type 2 inflammation (), enabling non-invasive preoperative inflammatory endotyping.

3.2 Radiomics and deep learning for endotyping

The evolution of CT-based AI studies is outlined in Tables 1A,C. Early CT-based artificial intelligence studies in CRSwNP relied on radiomics, which extracts handcrafted intensity, texture and spatial features from manually segmented CT regions (, ). In a multicenter study involving 431 patients from three institutions, Zhu et al. developed a radiomics model using 10 imaging features selected by least absolute shrinkage and selection operator (LASSO) regression, achieving an area under the receiver operating characteristic curve (AUC) of 0.815 for identifying eosinophilic CRSwNP, outperforming the conventional ethmoid-to-maxillary ratio model (AUC 0.655) (). Nevertheless, manual segmentation and laborious feature engineering restrict radiomic reproducibility and clinical scalability ().

These limitations have driven increasing interest in deep learning approaches that automatically learn hierarchical image representations directly from CT data. In one of the earliest endotyping studies, Du et al. developed a residual network-18 (ResNet-18) model using 29,993 CT slices from 251 CRSwNP patients (86 eosinophilic and 165 non-eosinophilic CRSwNP) and achieved a testing AUC of 0.963 for eosinophilic endotype prediction, demonstrating that CT images may encode information reflective of underlying inflammatory biology (). Building upon this concept, Lai et al. introduced a multichannel ResNet framework incorporating complementary imaging information from different CT planes, achieving an AUC of 0.984 in a cohort of 192 patients (). Subsequently, Zou et al. proposed a lightweight multi-view fusion network (ResMini) that maintained excellent predictive performance (AUC 0.991) while substantially reducing computational complexity, thereby improving the feasibility of real-world deployment (). More recently, Li et al. developed and externally validated a multicenter deep-learning model using 1,098 patients from two institutions. By integrating deep learning-derived CT features with clinical characteristics, the model achieved external AUCs exceeding 0.82 and demonstrated improved generalizability compared with previous single-center studies (). Compared with radiomics, deep-learning models automatically learn hierarchical CT representations and reduce dependence on predefined features. Multi-view or lightweight architectures may improve anatomic coverage and deployment feasibility, whereas CT-clinical models enhance clinical realism by integrating imaging signatures with blood or clinical variables. However, AUCs should be compared cautiously across studies because eCRS/eCRSwNP definitions, eosinophil thresholds, CT inputs, segmentation strategies, and validation designs vary substantially; externally or multicenter-validated models therefore provide stronger translational evidence than single-center models with very high internal performance.

3.3 Clinical translation and current limitations

For rhinologists, the greatest value of computed tomography-based artificial intelligence (CT-AI) lies not in replacing conventional radiologic scoring systems but in enabling non-invasive inflammatory endotyping before histopathologic information becomes available. It detects CT signatures linked to tissue eosinophilia and type 2 inflammation, bridging anatomical scans and endotype-directed evaluation to stratify recurrence risk and optimize therapy (, , ).

Despite promising results, current CT-based AI studies for CRSwNP endotyping have several important limitations. Most published studies are retrospective and single-center, limiting external validity and increasing the risk of selection bias (). In addition, definitions of eosinophilic CRS and type 2-high disease vary substantially across studies, with inconsistent thresholds for tissue eosinophilia and differing biomarker-based criteria (). Asian populations carry higher rates of non-type 2 and mixed inflammatory subtypes, limiting model transfer across geographic regions (–). Technical heterogeneity in CT scanners, slice thickness, reconstruction algorithms, and acquisition protocols also affects model robustness and reproducibility (). Finally, although many models report excellent discriminative performance, prospective clinical utility assessments remain lacking. A high AUC for predicting an endotype label should not be interpreted as evidence that a model can directly guide biologic selection, as studies demonstrating improvements in treatment decision-making and patient outcomes are still needed ().

4 Digital pathology and pathomics for spatial type 2 inflammatory signatures

4.1 The continuing role and limitations of conventional histopathology

CRSwNP offers a unique advantage: sinonasal tissue is routinely obtainable via surgery or biopsy, allowing histopathological assessment of eosinophilic infiltration, neutrophilic inflammation, basement membrane thickening, edema, goblet cell hyperplasia, glandular changes and fibrin deposition (). However, conventional pathology faces important limitations: subjective field selection, inconsistent cut-offs for eosinophil counts, spatial heterogeneity within the same polyp, unstable linkage to molecular endotypes, and the fact that routine hematoxylin and eosin (H&E) sections may contain spatial patterns that escape human recognition. These challenges create an unmet need for digital solutions that can standardize and quantify histopathological features in a spatially resolved manner.

4.2 Whole-slide imaging, pathomics, and molecularly informed endotyping

The evolution of AI-assisted digital pathology in CRSwNP is summarized in Tables 1B,C. Digital pathology combined with whole-slide image (WSI) analysis is transforming the evaluation of nasal polyp tissue from descriptive histology into quantitative, AI-driven endotyping. Early applications primarily focused on automating tasks traditionally performed by pathologists. The AI chronic rhinosinusitis evaluation platform (AICEP) initially enabled rapid and objective quantification of tissue eosinophilia from H&E slides, improving the efficiency and reproducibility of eosinophilic CRSwNP identification (). Subsequently, AICEP 2.0 expanded this approach beyond eosinophil counting by simultaneously recognizing eosinophils, neutrophils, lymphocytes, and plasma cells on WSIs, defining four distinct inflammatory subtypes associated with different recurrence risks and outperforming manual high-power field assessment (). More recently, the nasal polyps subtype diagnosis system (NPSS) framework further extended WSI-based analysis by integrating multidimensional cellular endotyping with three-dimensional reconstruction, enabling visualization of inflammatory cell spatial distributions and improving prognostic prediction for postoperative recurrence ().

Recent advances suggest that digital pathology can extend beyond cell enumeration toward molecular endotyping. The HE2Signature model linked H&E WSIs with transcriptomic endotype signatures using interpretable deep learning. Trained on 70 nasal polyp WSIs and externally validated in 224 patients from four centers, it identified T1, T2, and T3 endotypes with AUCs of 0.833, 0.903, and 0.935, respectively, and predicted the T2-associated FCER2/CST1 signature (). These findings indicate that routine H&E slides contain latent molecular and spatial information that AI can convert into clinically relevant biomarkers for precision rhinology.

Methodologically, these studies show that AI pathology in CRSwNP is moving from automated cell counting toward spatial, prognostic, and molecular inference. Early WSI models mainly reduced sampling bias in eosinophil quantification, whereas later systems expanded to multi-cellular phenotyping, inflammatory subtype classification, and spatial visualization of the tissue microenvironment. More recent models further suggest that routine H&E slides may contain latent prognostic or molecular information that can be extracted by interpretable deep learning.

4.3 Clinical translation and current limitations

From a rhinologist’s perspective, the greatest clinical value of digital pathology lies in postoperative inflammatory endotyping, as tissue specimens provide direct information on local inflammatory endotypes and tissue remodeling. AI-powered digital pathology extends conventional histopathological assessment by enabling objective quantification of tissue eosinophilia and inflammatory cell composition (), identifying spatial patterns within the inflammatory microenvironment (), and potentially linking routine H&E morphology with underlying molecular endotypes (). These advances position digital pathology as a promising tool for refining postoperative endotyping and recurrence risk stratification in CRSwNP.

Nevertheless, several barriers remain before these approaches can be integrated into routine clinical practice. Standardization of tissue sampling sites, section thickness, H&E staining protocols, scanner specifications, and image quality remains unresolved, while inter-institutional variability in pathology workflows may affect model generalizability (). Moreover, most AI-based analyses still require pathologist oversight for histological interpretation, validation, and clinical contextualization. Consequently, digital pathology should currently be regarded as a complement rather than a replacement for conventional pathological assessment.

5 Multimodal clinical machine learning

5.1 Why multimodal AI is more clinically realistic than single-modality AI

Current AI models in CRSwNP are predominantly developed within a single data modality, most commonly CT imaging, digital pathology, or blood biomarkers. While these approaches have advanced sinonasal endotyping, clinical decision-making in CRSwNP inherently relies on integrating information from multiple domains rather than interpreting each modality in isolation. In routine practice, rhinologists simultaneously consider symptoms, endoscopic findings, imaging characteristics, biomarkers, comorbidities, and treatment history when assessing inflammatory burden, estimating recurrence risk, and evaluating suitability for biologic therapy (, ).

Therefore, multimodal AI is not only a strategy to improve performance but a framework that mirrors rhinologists’ clinical reasoning. By integrating complementary data streams, it may bridge isolated biomarkers and real-world treatment stratification.

5.2 Toward an interpretable multimodal research framework

Published CRSwNP AI studies have begun to combine selected data sources, but the current evidence represents partial pairwise or multi-source integration rather than full cross-modal modeling (Table 1C). Examples include CT combined with clinical or blood variables (, ), imaging phenotypes explored alongside proteomic data (), and H&E morphology linked to transcriptomic labels (). These studies demonstrate the potential value of complementary information but do not establish an integrated CT–pathology model. To our knowledge, no published model has jointly integrated sinus CT, digital whole-slide pathology, and biomarkers for CRSwNP inflammatory endotyping. Figure 1 is therefore presented as a proposed future research architecture.

Any future implementation should prespecify modality-specific inputs, including CT radiomic or deep-learning features, histological cell-composition and spatial features, systemic or local biomarkers, and clinically relevant covariates. Feature weights should be learned within the development cohort rather than assigned a priori and should be locked before external validation. Reference standards should be task-specific, such as a prespecified consensus tissue-defined inflammatory endotype, standardized postoperative recurrence criteria, or validated biologic-response definitions. Missing inputs should be addressed using transparent imputation strategies, missingness indicators, or modality-aware modeling, accompanied by sensitivity analyses. Evaluation should include discrimination, calibration, incremental clinical utility, fairness, and prospective multicenter transportability.

6 AI-assisted biologic decision support and monitoring

6.1 The challenge of biologic target selection in CRSwNP

Biologic selection in CRSwNP is becoming increasingly challenging as therapeutic options now extend beyond anti-IL-4Rα to include anti-IgE, anti-IL-5/IL-5R, and emerging anti-thymic stromal lymphopoietin (TSLP) strategies (, ). However, patients with apparently similar type 2 disease often differ substantially in symptom burden, comorbidities, inflammatory profiles, recurrence risk, and treatment priorities. Moreover, validated predictive biomarkers capable of matching individual patients to specific biologic targets remain limited, and direct head-to-head comparisons between biologics are lacking. Consequently, biologic selection has evolved into a multidimensional decision-making process that requires integration of heterogeneous clinical information rather than reliance on a single biomarker or guideline criterion.

6.2 Predictive modeling for eligibility, response domains, and treatment trajectories

Emerging evidence suggests that AI may support several stages of biologic management in CRSwNP. Machine-learning approaches have demonstrated the ability to identify eosinophilic inflammation and type 2 inflammatory endotypes in CRSwNP (, , ). Because assessment of type 2 inflammatory burden is central to biologic decision-making, these approaches may facilitate identification of patients who warrant consideration for biologic therapy (). More recently, predictive models have been applied to biologic treatment datasets to estimate treatment response. In a real-world cohort of 84 CRSwNP patients receiving mepolizumab, machine-learning algorithms, particularly extreme gradient boosting (XGBoost), predicted sustained super-response over 24 months with an receiver operating characteristic area under the curve (ROC-AUC) of 0.766, identifying baseline blood eosinophilia, neutrophilia, and AERD comorbidity as important predictors of treatment outcomes (). These findings suggest that AI may not only enhance inflammatory endotyping but also support prediction of biologic response trajectories using multidimensional clinical data.

6.3 Large language models: guideline synthesis rather than autonomous selection

Large language models (LLMs) represent a distinct but rapidly evolving area of interest in biologic decision support. Recent studies evaluating ChatGPT for biologic selection in CRSwNP have reported moderate agreement with expert rhinology recommendations, although substantial variability was observed across different biologic agents (–). These findings suggest that LLMs may be useful for synthesizing guideline recommendations, summarizing evidence, and supporting clinician education. However, current evidence remains insufficient to support autonomous biologic selection, particularly given concerns regarding consistency, transparency, accountability, and the absence of patient-specific predictive capabilities. At present, LLMs should therefore be regarded as clinician-facing decision-support tools rather than substitutes for multidisciplinary clinical judgment.

7 Validation, explainability, and clinical governance

Despite encouraging performance in imaging, digital pathology, and biologic-response prediction, several challenges must be addressed before AI can be integrated into routine CRSwNP care. First, the lack of standardized reference labels remains a major barrier. Definitions of eosinophilic CRS, type 2-high disease, biologic eligibility, and treatment response vary substantially across studies, limiting comparability and model transportability (). Moreover, most published models are based on retrospective single-center cohorts, with relatively few undergoing rigorous external validation (37, 38). Future studies should prioritize prospective multicenter validation, calibration assessment, and demonstration of incremental clinical utility beyond existing clinical assessment tools (39). Particular attention should also be paid to biases arising from geographic variation in inflammatory endotypes, imaging protocols, pathology workflows, and patient selection, all of which may affect model fairness and generalizability (40).

AI systems must also be interpretable and compatible with existing clinical workflows. Explainable approaches, including feature importance analysis, attention maps, and biologically informed models, may improve clinician confidence and facilitate validation against established disease mechanisms (41). However, AI should support rather than replace clinician judgment, particularly for biologic selection (42). Robust data governance addressing privacy, transparency, accountability, and data sharing will be essential for responsible implementation (43). Ultimately, successful AI deployment will depend on clinically actionable information that integrates seamlessly into multidisciplinary care.

8 Discussion and conclusion

CRSwNP fulfills two principal features that make it a clinically tractable human model of mucosal type 2 inflammation. On one hand, many patients exhibit core molecular, cellular, and tissue-remodeling features of type 2 inflammatory disease, including epithelial alarmin signaling, activation of type 2 innate lymphoid cells and Th2 lymphocytes, tissue eosinophilia, local mucosal IgE production, and characteristic patterns of mucosal remodeling (, , –). Its frequent coexistence with asthma and N-ERD further enables investigation of type 2 inflammation across the united airway. On the other hand, CRSwNP offers substantial clinical research tractability. Polyp tissue obtained during clinically indicated surgery supports multi-omics and spatial pathological analyses, while standardized CT, endoscopy, and serial blood or nasal biomarkers enable longitudinal monitoring. Approved type 2-targeted biologics also permit prospective assessment of pathway-specific treatment responses. Nevertheless, CRSwNP is not a universal surrogate for all type 2 inflammatory disorders, given geographic variation and the presence of non-type 2 and mixed inflammatory endotypes.

The evidence reviewed here suggests that AI in CRSwNP is evolving from isolated classification tasks toward practical clinical decision support. Instead of pursuing high predictive AUC alone, modern AI generates interpretable assessments of inflammatory heterogeneity to guide endotyping, risk stratification and treatment design (42). This evolution mirrors rhinology’s shift from clinical phenotype-based diagnosis to precision medicine centered on treatable immune mechanisms.

First, CT-based AI may support non-invasive preoperative estimation of type 2-high endotypes and reduce reliance on postoperative tissue-based assessment. CRSwNP is well suited for CT-based AI research because sinus CT is commonly acquired in routine rhinology visits, delivering high-dimensional texture and spatial information invisible to conventional Lund-Mackay scoring (). Second, AI-assisted digital pathology may standardize H&E whole-slide inflammatory cell quantification post-surgery, reducing manual counting bias and revealing invisible spatial inflammatory microenvironment patterns. Third, next-generation multimodal AI requires three critical improvements: standardized multicenter datasets, cross-regional prospective validation, and embedded explainable artificial intelligence (XAI) modules for explainable cross-modal fusion. Existing multimodal models generally rely on partial feature-level integration and lack large prospective validation. CRSwNP therefore provides a clinically accessible research context for AI-assisted inflammatory endotyping. Routine rhinology practice readily yields clinical phenotypes, sinus CT images, surgical polyp specimens, systemic biomarkers and follow-up therapeutic data, a complete set of multimodal resources seldom accessible together for other inflammatory diseases. AI’s core advantage lies in converting fragmented multi-source clinical data into intuitive risk and stratification indicators to support rhinologists’ clinical judgment.

Future work should develop lightweight, externally validated multimodal pipelines compatible with hospital workflows, integrating CT, digital pathology, lab biomarkers and endoscopic data to achieve unified cross-center inflammatory endotyping.

Author contributions

XJ: Writing – original draft, Writing – review & editing. TK: Writing – original draft, Writing – review & editing. CC: Writing – original draft. HQ: Writing – original draft. WL: Funding acquisition, Investigation, Writing – original draft, Writing – review & editing. AS: Funding acquisition, Supervision, Writing – original draft, Writing – review & editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This study was supported by National High Level Hospital Clinical Research Funding (Grant No. 2025-PUMCH-C-031); CAMS Innovation Fund for Medical Sciences (Grant No. CIFMS-2024-I2M-C&T-B-030).

Acknowledgments

We gratefully acknowledge the financial support from National High Level Hospital Clinical Research Funding (Grant No. 2025-PUMCH-C-031); CAMS Innovation Fund for Medical Sciences (Grant No. CIFMS-2024-I2M-C&T-B-030). Additionally, Figure 1 was created by biorender.com.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Abbreviations

AERD, aspirin-exacerbated respiratory disease; AI, artificial intelligence; AICEP, AI chronic rhinosinusitis evaluation platform; AUC, area under the curve; CNN, convolutional neural network; CRS, chronic rhinosinusitis; eCRS, eosinophilic chronic rhinosinusitis; CRSwNP, chronic rhinosinusitis with nasal polyps; CT, computed tomography; CT-AI, computed tomography-based artificial intelligence; DL, deep learning; ESS, endoscopic sinus surgery; FeNO, fractional exhaled nitric oxide; H&E, hematoxylin and eosin; IgE, immunoglobulin E; IL, interleukin; IL-4Rα, interleukin-4 receptor alpha; IL-5R, interleukin-5 receptor; LASSO, least absolute shrinkage and selection operator; LLM, large language model; ML, machine learning; N-ERD, NSAID-exacerbated respiratory disease; N/A, not applicable; NS, not specified; NSAID, nonsteroidal anti-inflammatory drug; PROM, patient-reported outcome measure; ROC, receiver operating characteristic; ROC-AUC, receiver operating characteristic area under the curve; ResMini, lightweight multi-view fusion network; ResNet, residual network; ResNet-18, residual network-18; SNOT-22, Sino-Nasal Outcome Test-22; TSLP, thymic stromal lymphopoietin; WSI, whole-slide image; XAI, Explainable Artificial Intelligence; XGBoost, extreme gradient boosting.

References

• 1.

  KatoASchleimerRPBleierBS. Mechanisms and pathogenesis of chronic rhinosinusitis. J Allergy Clin Immunol. (2022) 149(5):1491–503. 10.1016/j.jaci.2022.02.016

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2.

  SchleimerRP. Immunopathogenesis of chronic rhinosinusitis and nasal polyposis. Annu Rev Pathol. (2017) 12:331–57. 10.1146/annurev-pathol-052016-100401

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3.

  BachertCZhangNCavaliereCWeipingWGevaertEKryskoO. Biologics for chronic rhinosinusitis with nasal polyps. J Allergy Clin Immunol. (2020) 145(3):725–39. 10.1016/j.jaci.2020.01.020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4.

  FokkensWJViskensA-SBackerVContiDDe CorsoEGevaertPet al. EPOS/EUFOREA update on indication and evaluation of biologics in chronic rhinosinusitis with nasal polyps 2023. Rhinology. (2023) 61(3):194–202. 10.4193/Rhin22.489

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5.

  TomassenPVandeplasGVan ZeleTCardellL-OArebroJOlzeHet al. Inflammatory endotypes of chronic rhinosinusitis based on cluster analysis of biomarkers. J Allergy Clin Immunol. (2016) 137(5):1449–56.e4. 10.1016/j.jaci.2015.12.1324

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6.

  WangWXuYWangLZhuZAodengSChenHet al. Single-cell profiling identifies mechanisms of inflammatory heterogeneity in chronic rhinosinusitis. Nat Immunol. (2022) 23(10):1484–94. 10.1038/s41590-022-01312-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7.

  LiaoGNakayamaTZhuBLeeITYeungJYeoYYet al. Multi-scaled transcriptomics of chronically inflamed nasal epithelium reveals immune-epithelial dynamics and tissue remodeling in nasal polyp formation. Immunity. (2025) 58(10):2593–608.e6. 10.1016/j.immuni.2025.08.009

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8.

  NazariJShahbaFJafariaghdamNMohebbiSArshiSBemanianMHet al. Immune endotyping and gene expression profile of patients with chronic rhinosinusitis with nasal polyps in the aspirin-exacerbated respiratory disease (AERD) and the non-AERD subgroups. Allergy Asthma Clin Immunol. (2024) 20(1):14. 10.1186/s13223-024-00876-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9.

  BachertCMarpleBSchlosserRJHopkinsCSchleimerRPLambrechtBNet al. Adult chronic rhinosinusitis. Nat Rev Dis Primers. (2020) 6(1):86. 10.1038/s41572-020-00218-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10.

  FokkensWJLundVJHopkinsCHellingsPWKernRReitsmaSet al. European Position paper on rhinosinusitis and nasal polyps 2020. Rhinology. (2020) 58(Suppl S29):1–464. 10.4193/Rhin20.600

  • CrossRef
  • Google Scholar
  • View reference in article

• 11.

  LundVJMackayIS. Staging in rhinosinusitus. Rhinology. (1993) 31(4):183–4. PMID: 8140385

  • Pubmed Abstract
  • Google Scholar
  • View reference in article

• 12.

  MasseyCJRamosLBeswickDMRamakrishnanVRHumphriesSM. Clinical validation and extension of an automated, deep learning-based algorithm for quantitative Sinus CT analysis. AJNR Am J Neuroradiol. (2022) 43(9):1318–24. 10.3174/ajnr.A7616

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13.

  GilliesRJKinahanPEHricakH. Radiomics: images are more than pictures, they are data. Radiology. (2016) 278(2):563–77. 10.1148/radiol.2015151169

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14.

  LambinPLeijenaarRTHDeistTMPeerlingsJDe JongEECVan TimmerenJet al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. (2017) 14(12):749–62. 10.1038/nrclinonc.2017.141

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15.

  ZhuK-ZHeCLiZWangP-JWenS-XWenK-Xet al. Development and multicenter validation of a novel radiomics-based model for identifying eosinophilic chronic rhinosinusitis with nasal polyps. Rhinology. (2023) 61(2):132–43. 10.4193/Rhin22.361

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16.

  Darbari KaulRSacksP-LThielCRimmerJKalishLCampbellRGet al. Radiomics of the paranasal sinuses: a systematic review of computer-assisted techniques to assess computed tomography radiological data. Am J Rhinol Allergy. (2025) 39(2):147–58. 10.1177/19458924241304082

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17.

  DuWKangWLaiSCaiZChenYZhangXet al. Deep learning in computed tomography to predict endotype in chronic rhinosinusitis with nasal polyps. BMC Med Imaging. (2024) 24(1):25. 10.1186/s12880-024-01203-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18.

  LaiSKangWChenYZouJWangSZhangXet al. An End-to-End CRSwNP prediction with multichannel ResNet on computed tomography. Int J Biomed Imaging. (2024) 2024:4960630. 10.1155/2024/4960630

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19.

  ZouJLyuYLinYChenYLaiSWangSet al. A multi-view fusion lightweight network for CRSwNPs prediction on CT images. BMC Med Imaging. (2024) 24(1):112. 10.1186/s12880-024-01296-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20.

  LiJMaoNAodengSZhangHZhuZWangLet al. Development and validation an integrated deep learning model to assist eosinophilic chronic rhinosinusitis diagnosis: a multicenter study. Int Forum Allergy Rhinol. (2025) 15(10):1070–81. 10.1002/alr.23595

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21.

  PetsiouDPSpinosDMartinosAMuzaffarJGarasGGeorgalasC. Effectiveness of artificial intelligence in detecting sinonasal pathology using clinical imaging modalities: a systematic review. Rhinology. (2025) 63(4):448–62. 10.4193/Rhin25.044

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22.

  WangXZhangNBoMHoltappelsGZhengMLouHet al. Diversity of T(H) cytokine profiles in patients with chronic rhinosinusitis: a multicenter study in Europe, Asia, and oceania. J Allergy Clin Immunol. (2016) 138(5):1344–53. 10.1016/j.jaci.2016.05.041

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23.

  MinJ-YKimJYSungCMKimSTChoH-JMunSJet al. Inflammatory endotypes of chronic rhinosinusitis in the Korean population: distinct expression of type 3 inflammation. Allergy Asthma Immunol Res. (2023) 15(4):437–50. 10.4168/aair.2023.15.4.437

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24.

  LouHZhangNBachertCZhangL. Highlights of eosinophilic chronic rhinosinusitis with nasal polyps in definition, prognosis, and advancement. Int Forum Allergy Rhinol. (2018) 8(11):1218–25. 10.1002/alr.22214

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25.

  ZwanenburgAVallièresMAbdalahMAAertsHJWLAndrearczykVApteAet al. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology. (2020) 295(2):328–38. 10.1148/radiol.2020191145

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26.

  WuQChenJDengHRenYSunYWangWet al. Expert-level diagnosis of nasal polyps using deep learning on whole-slide imaging. J Allergy Clin Immunol. (2020) 145(2):698–701.e6. 10.1016/j.jaci.2019.12.002

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27.

  WuQChenJRenYQiuHYuanLDengHet al. Artificial intelligence for cellular phenotyping diagnosis of nasal polyps by whole-slide imaging. EBioMedicine. (2021) 66:103336. 10.1016/j.ebiom.2021.103336

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28.

  LuoXLiHChenJZhouZLiangGChaoMet al. Artificial intelligence-based pathological subtype diagnosis of nasal polyps: a multidimensional and micro-visualization study. Allergy. (2026) 81(5):1665–79. 10.1111/all.70282

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29.

  WangKRenYMaLHanLXuLGuanHet al. Explainable artificial intelligence predicts inflammatory and spatial heterogeneity from nasal polyp histology. J Allergy Clin Immunol. (2025) 156(6):1573–87. 10.1016/j.jaci.2025.08.016

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30.

  BeraKSchalperKARimmDLVelchetiVMadabhushiA. Artificial intelligence in digital pathology—new tools for diagnosis and precision oncology. Nat Rev Clin Oncol. (2019) 16(11):703–15. 10.1038/s41571-019-0252-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31.

  XiongPChenJZhangYShuLShenYGuYet al. Predictive modeling for eosinophilic chronic rhinosinusitis: nomogram and four machine learning approaches. iScience. (2024) 27(2):108928. 10.1016/j.isci.2024.108928

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32.

  Domínguez-SosaMSCabrera-RamírezMSMarrero-RamosMCCabrera-LópezCCarrillo-DíazTBenítez-RosarioJet al. Using machine learning to define mepolizumab treatment response at 2 years in patients with chronic rhinosinusitis with nasal polyps. Front Allergy. (2026) 7:1710163. 10.3389/falgy.2026.1710163

  • CrossRef
  • Google Scholar
  • View reference in article

• 33.

  ManzellaRImmordinoAGallettiCGiammona IndacoFStiloGMessinaGet al. ChatGPT in the management of chronic rhinosinusitis with nasal polyps: promising support or digital illusion? Insights from a multicenter observational study. J Clin Med. (2025) 14(13):4501. 10.3390/jcm14134501

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34.

  SireciFLorussoFImmordinoACentineoMGerardiIPattiGet al. ChatGPT as a new tool to select a biological for chronic rhino sinusitis with polyps, “caution advised” or “distant reality”?J Pers Med. (2024) 14(6):563. 10.3390/jpm14060563

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35.

  ManiaciASaibeneAMCalvo-HenriquezCVairaLRadulescoTMichelJet al. Is generative pre-trained transformer artificial intelligence (Chat-GPT) a reliable tool for guidelines synthesis? A preliminary evaluation for biologic CRSwNP therapy. Eur Arch Otorhinolaryngol. (2024) 281(4):2167–73. 10.1007/s00405-024-08464-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 36.

  CollinsGSMoonsKGMDhimanPRileyRDBeamALVan CalsterBet al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. Br Med J. (2024) 385:e078378. 10.1136/bmj-2023-078378

  • CrossRef
  • Google Scholar
  • View reference in article

• 37.

  LiuXCruz RiveraSMoherDCalvertMJDennistonAKChanA-Wet al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med. (2020) 26(9):1364–74. 10.1038/s41591-020-1034-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38.

  Cruz RiveraSLiuXChanA-WDennistonAKCalvertMJDarziAet al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat Med. (2020) 26(9):1351–63. 10.1038/s41591-020-1037-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39.

  Van CalsterBMcLernonDJVan SmedenMWynantsLSteyerbergEW. Calibration: the achilles heel of predictive analytics. BMC Med. (2019) 17(1):230. 10.1186/s12916-019-1466-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40.

  KellyCJKarthikesalingamASuleymanMCorradoGKingD. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. (2019) 17(1):195. 10.1186/s12916-019-1426-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 41.

  CutilloCMSharmaKRFoschiniLKunduSMackintoshMMandlKDet al. Machine intelligence in healthcare-perspectives on trustworthiness, explainability, usability, and transparency. NPJ Digit Med. (2020) 3:47. 10.1038/s41746-020-0254-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42.

  TopolEJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. (2019) 25(1):44–56. 10.1038/s41591-018-0300-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43.

  OrganizationWH. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. Geneva: World Health Organization (2021).

  • Google Scholar

• 44.

  WangYLiuFZhangHWangQYuPWangJet al. Deep learning model for the differential diagnosis of nasal polyps and inverted papilloma by CT images: a multicenter study. Acad Radiol. (2025) 32(5):2900–9. 10.1016/j.acra.2024.12.011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45.

  GuoMZangXFuWYanHBaoXLiTet al. Classification of nasal polyps and inverted papillomas using CT-based radiomics. Insights Imaging. (2023) 14(1):188. 10.1186/s13244-023-01536-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46.

  HuaH-LLiSXuYChenS-KongY-YangRet al. Differentiation of eosinophilic and non-eosinophilic chronic rhinosinusitis on preoperative computed tomography using deep learning [J]. Clin Otolaryngol. (2023) 48(2):330–8. 10.1111/coa.13988

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47.

  HumphriesSMCentenoJPNotaryAMGerowJCicchettiGKatialRKet al. Volumetric assessment of paranasal sinus opacification on computed tomography can be automated using a convolutional neural network. Int Forum Allergy Rhinol. (2020) 10(11):1218–25. 10.1002/alr.22588

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 48.

  ShenZWeiYLiuKMaZZhangZWangXet al. Deep learning-derived quantitative scores for chronic rhinosinusitis assessment: correlation with quality of life outcomes. Am J Rhinol Allergy. (2025) 39(3):187–96. 10.1177/19458924251313845

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49.

  MasseyCJHumphriesSMMaceJCSmithTLSolerZMRamakrishnanVR. Multi-institutional validation of an AI-based sinus CT analytic platform with olfactory assessments. Int Forum Allergy Rhinol. (2024) 14(11):1806–9. 10.1002/alr.23410

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 50.

  HsuYCLinKTLeeMSShenLSYehTHLinYT. Multiple instance learning for eosinophil quantification of sinonasal histopathology images: a hierarchical determination on whole slide images. Int Forum Allergy Rhinol. (2024) 14(9):1513–6. 10.1002/alr.23365

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51.

  LiuWLiuXZhangNLiJWenYWeiYet al. Microvessel quantification by fully convolutional neural networks associated with type 2 inflammation in chronic rhinosinusitis. Ann Allergy Asthma Immunol. (2022) 128(6):697–704.e1. 10.1016/j.anai.2022.02.025

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52.

  WangKRenYMaLFanYYangZYangQet al. Deep learning-based prediction of treatment prognosis from nasal polyp histology slides [J]. Int Forum Allergy Rhinol. (2023) 13(5):886–98. 10.1002/alr.23083

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 53.

  RaghavanAMAboueishaMAProhnitchiICvancaraDJHumphreysIMJafariAet al. Using machine learning models to diagnose chronic rhinosinusitis: analysis of Pre-treatment patient-generated health data to predict cardinal symptoms and sinonasal inflammation. Am J Rhinol Allergy. (2025) 39(3):229–36. 10.1177/19458924251322081

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 54.

  IshikawaMJiangZNguyenCHHatsukawaHMamitsukaH. Clinical key features uncovered by blood eosinophilia-based machine learning classification of chronic rhinosinusitis. Int Forum Allergy Rhinol. (2026) 16(4):353–68. 10.1002/alr.70077

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 55.

  ThorwarthRMScottDWLalDMarinoMJ. Machine learning of biomarkers and clinical observation to predict eosinophilic chronic rhinosinusitis: a pilot study. Int Forum Allergy Rhinol. (2021) 11(1):8–15. 10.1002/alr.22632

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56.

  ZhouHFanWQinDLiuPGaoZLvHet al. Development, validation and comparison of artificial neural network and logistic regression models predicting eosinophilic chronic rhinosinusitis with nasal polyps [J]. Allergy Asthma Immunol Res. (2023) 15(1):67–82. 10.4168/aair.2023.15.1.67

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 57.

  MorgensternCBartosikTJBayerKNCampionNJFrommletFGanglKet al. Proteomic profiling and machine learning for endotype prediction in chronic rhinosinusitis. J Allergy Clin Immunol. (2026) 157(1):190–202. 10.1016/j.jaci.2025.08.025

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58.

  Moreno-JiménezEMorgadoNGarcía-SánchezATriviñoJCEstravísMGómez-GarcíaMet al. Unsupervised machine-learning-based endotype discovery using iterative resampling in dupilumab-treated patients. Int J Mol Sci. (2026) 27(12):5266. 10.3390/ijms27125266

  • CrossRef
  • Google Scholar

• 59.

  HeSChenWWangXXieXLiuFMaXet al. Deep learning radiomics-based preoperative prediction of recurrence in chronic rhinosinusitis. iScience. (2023) 26(4):106527. 10.1016/j.isci.2023.106527

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

Keywords

artificial intelligence, biologics, chronic rhinosinusitis with nasal polyps, deep learning, digital pathology, multimodal modeling, radiomics, type 2 inflammation

Citation

Jiang X, Kang T, Chen C, Qiao H, Lv W and Surita A (2026) Artificial intelligence-assisted type 2 inflammatory endotyping in CRSwNP: from Sinus CT and digital pathology to biologic decision support. Front. Allergy 7:1923079. doi: 10.3389/falgy.2026.1923079

Updates

Copyright

*Correspondence: Wei Lv Aodeng Surita

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.