Leveraging large language models to enhance cytopathology: Opportunities, challenges, and future directions; a practical review from the ASC Clinical Practice Committee

Abstract

Large language models (LLMs) and vision‐language models represent a fundamentally different category of artificial intelligence (AI) compared to prior image analysis approaches in digital pathology, which have largely been based on convolutional neural network architectures. This review from the American Society of Cytopathology Clinical Practice Committee examines the current evidence for LLM and vision‐language model applications in cytopathology, including structured reporting, diagnostic assistance, quality control, education, and workflow integration. The distinction between applications with preliminary evidence and those that remain hypothetical is described. A detailed assessment of the challenges that must be addressed before clinical deployment, including hallucination risk, limited explainability, bias, data privacy, validation gaps, and infrastructure barriers is discussed. A review of the regulatory landscape in the United States and European Union as it applies to AI‐enabled software as a medical device is provided. Recommendations addressing cytopathology‐specific benchmarks, multi‐institutional validation, transparent governance, and incremental deployment beginning with low‐risk applications are suggested. In the current environment, LLMs have the potential to augment cytopathology practice, but responsible adoption requires rigorous validation and sustained collaboration among cytopathologists, AI researchers, and regulatory bodies.

INTRODUCTION

The AI landscape: Where large language models fit

Artificial intelligence (AI) encompasses a family of computational methods whose distinctions matter for clinical application, regulation, and limitations. Machine learning (ML) is a subset of AI in which systems learn patterns from data; deep learning, a subset of machine learning, uses multilayered neural networks to learn representations directly from raw input. ¹ Three deep‐learning architectures are relevant to pathology. Convolutional neural networks (CNNs) extract spatial features from grid‐structured input and have been applied to image‐based tasks such as cell detection, lesion classification, and biomarker prediction ² , ³ , ⁴ ; the architecture also generalizes to non–image matrix‐structured laboratory data, but CNNs do not process natural language. Large language models (LLMs) are based on the transformer architecture ⁵ and are trained on large text corpora; their self‐attention mechanism captures contextual relationships across input sequences and supports tasks such as summarization, information extraction, question answering, and text generation. Generative pretrained transformers (GPTs) are one family of LLMs; the broader category of generative AI refers to any model that produces novel content. Vision‐language models (VLMs) extend LLMs by coupling an image encoder to a language model through a multimodal projector, ⁶ enabling joint image‐and‐text input.

This distinction among CNNs (image only), LLMs (text only), and VLMs (multimodal) has direct implications for validation, regulation, and deployment. In current commercial products, the distinction is increasingly blurred. Frontier general‐purpose models (e.g., GPT‐4o, Claude, Gemini) accept image and text natively, whereas pathology‐specific VLMs such as PathChat, BiomedCLIP, and Pathology Language‐Image Pretraining differ primarily in the composition and scale of their pathology training data rather than in core architecture. This review focuses on LLMs and their multimodal extensions, not on traditional CNN‐based image analysis.

AI in cytopathology: historical context

AI‐assisted screening in cytopathology predates the current era of large language models and has progressed through three broadly distinct eras. The first generation comprised rule‐based image‐analysis systems, such as FocalPoint GS and the ThinPrep Imaging System, which used classical morphometry to prescreen Pap tests and flag regions of interest. ⁷ , ⁸ The second generation introduced CNN‐based deep learning, exemplified by the Hologic Genius Digital Diagnostics system for cervical cytology ⁹ and more recent models for urine cytology, fine‐needle aspiration, and effusion specimens. ¹⁰ , ¹¹ In this review, “traditional AI” refers to both generations: image‐classification systems built on rule‐based morphometry or CNN architectures that produce categorical outputs without natural‐language capability. The third generation, the subject of this review, comprises LLMs and VLMs.

Compared with traditional AI, LLMs and VLMs add capabilities in summarization and structured extraction from free‐text reports, generation of standardized report content, synthesis of clinical data across documents, and text‐based interactive response to queries. We use these descriptions in an operational sense; LLMs generate statistically likely token sequences and should not be characterized as “understanding” content. Commercial LLMs are also not trained in real time; model behavior is frozen at the release of each version, and capability changes accompany scheduled updates rather than ongoing learning. This review examines these capabilities, their current evidence base, and the challenges that must be addressed before clinical adoption.

Methodology and scope

This is a narrative review prepared by compiling and reviewing relevant literature published between 2018 and 2025, with an emphasis on LLM‐specific work from 2022 onward. Sources were identified through PubMed and Google Scholar searches combining cytopathology and pathology terms (e.g., cytopathology, cytology, fine‐needle aspiration, Pap, Bethesda, Paris, Milan) with AI and NLP terms (e.g., LLM, VLM, GPT, ChatGPT, foundation model, natural language processing, transformer), supplemented by authors’ knowledge of the field. Selected conference abstracts (USCAP, ASC, CAP) if they included the most current available data on a specific application; their non–peer‐reviewed status is identified in the text. The scope is limited to applications to LLMs, VLMs, and natural‐language‐processing tools that applies to cytopathology and closely related anatomic pathology contexts.

Generative AI tools were used in a limited capacity for language editing and proofreading of author‐drafted text. All content was reviewed, verified, and approved by the authors, who take full responsibility for the work. AI tools were not used for literature search, study selection, data extraction, or generation of original scientific content.

Based on this review of the literature and expert committee discussion, recommendations regarding LLM use in cytopathology were developed and are summarized in Section 1.4 and detailed in Section 6.

Summary of recommendations

The committee’s recommendations are previewed here as a roadmap; each is treated with supporting evidence and rationale in Section 6: (1) anchor implementation in patient and practitioner well‐being; (2) develop cytopathology‐specific benchmarks and validation frameworks; (3) prioritize explainability and uncertainty quantification; (4) build multi‐institutional datasets while recognizing local fitting may be preferable for institution‐specific tools; (5) begin with low‐risk, high‐value applications; (6) establish governance and monitoring infrastructure; (7) integrate AI literacy into cytopathology training; and (8) engage with regulatory and professional organizations.

CURRENT LLM AND VLM APPLICATIONS IN CYTOPATHOLOGY

NLP and structured reporting

A strength of cytopathology is that diagnostic reporting relies on widely accepted standardized classification systems, including the Bethesda System for cervical and thyroid cytology, the Paris System for urinary cytology, and the Milan System for salivary gland cytopathology. ¹² , ¹³ , ¹⁴ Most laboratories use these classifications with standardized templates internally, but reporting style and the surrounding free‐text content still vary across institutions and individual practitioners, particularly when reports are aggregated across centers for registry, research, or quality assurance purposes.

LLMs have demonstrated the ability to extract structured data from free‐text pathology reports with high accuracy. Grothey et al. ¹⁵ showed that GPT‐4 can extract structured diagnostic and staging information from unstructured histopathology reports; Truhn et al. ¹⁶ reported similar performance in converting narrative surgical pathology reports into structured formats. In cytopathology specifically, Nandish et al. ¹⁷ trained random‐forest and neural‐network classifiers on 2879 free‐text breast fine‐needle aspiration (FNA) reports from a single institution, mapping 13 lesion types into five diagnostic categories with 99.4% test‐set accuracy and a reported histopathologic correlation of 98.6%. Compared with ICD‐ or SNOMED‐coded case search, an LLM can retrieve cases on the basis of free‐text morphologic descriptors not captured by structured codes (e.g., Hurthle cell change with microfollicular architecture), although it cannot recover findings that were not documented in the first place.

Beyond extraction, LLMs can support report generation by mapping narrative descriptions to standardized terminology; for example, recognizing that “follicular cells with nuclear atypia and pronounced Hurthle cell change” and “follicular cells with cytologic atypia and diffuse oncocytic change” describe equivalent findings and classifying both under the appropriate Bethesda category. LLMs can also map diagnostic codes such as SNOMED‐CT, which is essential for interoperability and downstream data analytics. ¹⁸ Structured outputs facilitate integration with cancer registries, clinical trials, and quality assurance programs.

These capabilities are promising but remain largely validated in surgical pathology rather than cytopathology. Further evaluation on cytology‐specific report formats, classification systems, and edge cases is needed before clinical deployment.

Diagnostic assistance and clinical decision support

LLM‐based systems have shown early potential as diagnostic consultation tools. AMIE (Articulate Medical Intelligence Explorer), an LLM‐based dialogue system, achieved 59.1% top‐10 diagnostic accuracy (the proportion of cases in which the correct diagnosis appeared in the model’s top 10 ranked differential diagnoses) compared with 33.6% for unassisted clinicians in a structured clinical evaluation. ¹⁹ Although this study was not cytopathology‐specific, it demonstrates the capacity of LLMs to generate differential diagnoses through multiturn clinical dialogue.

In pathology, clinical data summarization represents a practical near‐term application. Pathologists routinely review clinical notes, radiology reports, and prior pathology findings before signing out. LLMs can synthesize this information and present relevant clinical context alongside the specimen under review. ²⁰ For cytology, where interpretation often depends on clinical history (e.g., prior malignancy, imaging findings, molecular results), automated summarization could reduce the time spent navigating electronic health records and improve contextual interpretation.

LLMs can also assist with ancillary test selection, particularly in cytology where sample material is often limited. By integrating clinical and morphologic data, these tools could help prioritize which ancillary studies are most likely to yield actionable information. ¹⁹ , ²¹ In any such deployment the cytopathologist is the primary end user; downstream clinicians and patients see LLM‐influenced output only through the signed‐out report, for which the cytopathologist retains responsibility and liability. Static text warnings about hallucination risk are unlikely to influence behavior in isolation; active design measures (provenance tagging of AI‐generated text, audit logging, calibrated confidence display, and mandatory review steps) are more likely to be effective.

Vision‐language models in pathology

VLMs combine image encoders (typically vision transformers) with LLM components through multimodal projectors, enabling simultaneous processing of visual and textual data. ⁶ In pathology, two systems have been described as they apply to pathology‐related tasks with robust evaluation.

PathChat, developed by Lu et al., ⁶ is a VLM‐based diagnostic assistant that integrates histopathology and cytology images with clinical context. In diagnostic evaluations, PathChat achieved 78.1% accuracy with image input alone and 89.5% when clinical context was provided and was preferred over GPT‐4V by pathologist reviewers 56.5% of the time. This level of accuracy is well below the threshold for autonomous diagnostic use; the realistic near‐term role is as a consultation or second‐opinion tool, and the gain from added clinical context is itself a practical limitation given that comprehensive history is often unavailable at the time of cytologic interpretation.

BiomedCLIP, a multimodal foundation model trained on 15 million biomedical image‐text pairs, supports cross‐modal retrieval: given an image, it can retrieve relevant text descriptions and vice versa. ²² It achieved greater than 56% top 1 and 77% top 5 retrieval accuracy, with applications in education, case‐based search, and diagnostic support. Pathology Language‐Image Pretraining provides a complementary pathology‐specific dataset of more than 208,000 image‐caption pairs across 32 pathology subspecialties. ²³ A representative selection of these models is summarized in Table 1.

TABLE 1.

LLM and VLM systems evaluated in pathology and cytopathology.

System / study	Modality	Functionality	Performance	Cytopathology relevance
AMIE 19	McDuff et al. 2025Text (LLM)	Multiturn diagnostic dialogue and differential diagnosis generation	59.1% top 10 diagnostic accuracy vs 33.6% unassisted clinicians	Interactive diagnostic consultation
Nandish et al. 17	Text (NLP)	Free‐text cytopathology report classification using domain‐specific lexicons	$>$ 99% breast lesion classification accuracy; 98.6% histopathologic correlation	Direct cytopathology application: report standardization
Grothey et al. 15	Text (LLM)	Structured data extraction from unstructured pathology reports	High accuracy across multiple data elements	Report automation and structured data generation
PathChat 6	Multimodal (VLM)	Vision‐language diagnostic assistant with clinical context integration	78.1% image‐only; 89.5% with context; preferred over GPT‐4V 56.5% of time	Multimodal diagnostic reasoning
BiomedCLIP 22	Multimodal (VLM)	Cross‐modal retrieval from 15M biomedical image‐text pairs	$>$ 56% top‐1; 77% top‐5 retrieval accuracy	Image‐text search for education and diagnostics
Pathology Language‐Image Pretraining 23	Multimodal (VLM)	Pathology‐specific image‐language pretraining (208K + pairs, 32 subspecialties)	Top‐5 image‐to‐image retrieval accuracy 0.96–1.00 across Kather colon (9 classes), PanNuke (19 classes), and KIMIA Path24C (24 classes); modest improvement over baseline CLIP	Pathology‐specific multimodal pretraining and retrieval

Note: Systems are categorized by modality to distinguish text‐based and image‐based metrics. “Top 10 diagnostic accuracy” (AMIE) denotes the proportion of cases in which the correct diagnosis appeared in the model’s top ten ranked differentials. Reported figures are drawn from the cited primary sources and reflect the cohorts and benchmarks used in those studies; cross‐study comparison should therefore be made with caution.

A potential advantage of VLMs is the ability to return a textual rationale alongside a classification; for example, listing morphologic features cited in support of a diagnosis in response to a question such as “What features support adenocarcinoma over reactive mesothelial cells in this effusion specimen?” ²⁴ These rationales are produced by the same statistical token‐prediction process that produced the classification itself and should be regarded as candidate hypotheses to verify, not as faithful traces of the model’s internal computation. Even with this caveat, this contrasts with traditional CNN classifiers, which provide a label or probability without any explanation.

These systems are not yet deployed in clinical cytopathology practice. Their validation has been conducted primarily on histopathology data, and performance on cytology‐specific tasks (e.g., FNA interpretation, liquid‐based preparations) requires dedicated evaluation. Given the wide morphologic differences in various cytology preparations, direct application of these systems to cytopathology samples may not yield as a high a performance. The stochastic nature of LLM outputs and the risk of hallucination (discussed in Section 4) apply equally to VLMs. However, both PathChat and BiomedCLIP highlight the potential to have such an assistive technology purposely built for pathology.

Quality control and error detection

LLMs are well suited to quality control tasks in cytopathology, where report accuracy, completeness, and consistency are essential. Preliminary work presented at the 2025 USCAP Annual Meeting demonstrated that an LLM could identify 84% of errors that had previously led to amended reports, including laterality errors, narrative‐to‐synoptic inconsistencies, and incomplete reports, by cross‐referencing procedure notes, requisitions, and radiology reports. ²⁵ Saluja et al. ²⁶ further showed that LLMs can extract cancer type, stage, and prognostic data from structured pathology reports with high accuracy. Currently, quality control (QC) assessment is mostly performed manually or with minimal computer assistance, making tasks such as identifying reporting errors, performing cyto‐histo correlation, and computing ASCUS/SIL ratios time‐consuming.

In cytopathology, several QC applications are feasible. LLMs could review draft reports to detect inconsistencies between diagnostic text, specimen source, and ancillary test results. For example, in thyroid FNA cases classified as atypia of undetermined significance or follicular neoplasm, concurrent molecular testing results (reported separately) could be automatically cross‐referenced. An LLM could flag cases where molecular findings such as PTH gene expression positivity suggest a parathyroid lesion rather than an indeterminate thyroid nodule, a distinction with significant clinical management implications. ²⁷ Additional potential applications include automated cytology‐histology correlation, detection of missing adequacy statements, and case triage for targeted peer review based on discrepancy likelihood.

This evidence is preliminary. The work by Carmichael et al. has not yet undergone peer review and the conference abstract does not report a false‐positive rate (the proportion of LLM‐flagged items that on adjudication were not in fact errors), which is a critical metric for any QC tool because each false alarm consumes pathologist review time. It is also important to note that many laboratory QC tasks (laterality checks, presence of required fields, ASCUS/SIL ratios) are well handled by deterministic rule‐based middleware; the case for LLMs is strongest where the underlying logic depends on the semantic content of narrative text, and hybrid systems combining rule‐based gating with LLM review of flagged cases are likely the most defensible near‐term architecture. No prospective validation of LLM‐based QC in cytopathology has been reported, and implementation will also require secure integration with laboratory information systems, Health Insurance Portability and Accountability Act (HIPAA)–compliant data handling, and compatibility with diverse institutional reporting templates.

Emerging applications

The applications described in Section 2 have at least preliminary evidence supporting their feasibility. This section addresses areas where LLMs show potential but where cytopathology‐specific evidence remains limited or absent. Claims in this section should be understood as forward‐looking potential applications rather than validated.

Education and training

LLMs offer several applications in cytopathology education. When prompted with structured queries, applications such as ChatGPT can generate case‐based learning modules, differential diagnosis lists with associated immunostains and molecular markers, and multiple‐choice assessment questions. ²⁴ Faculty have begun adopting hybrid teaching strategies that combine traditional instruction with LLM‐assisted content generation to improve standardization across training programs. ²⁸ Trainee interest in AI‐assisted learning is broadly observed, though the durability of these tools in formal training programs remains to be evaluated.

For learners, LLMs can function as on‐demand tutoring systems, providing individualized feedback, self‐assessment tools, and literature summaries. When paired with VLMs, these systems can also assist in interpreting histopathology and cytology images, offering explanations alongside visual data. ²⁸ These capabilities could help address variability in training quality across programs, particularly at institutions with limited subspecialty faculty.

However, several limitations must be acknowledged. LLM‐generated educational content is susceptible to hallucination and factual errors. Faculty review is necessary but not sufficient. Hallucinations are an intrinsic property of how LLMs generate text where even material reviewed by a person can drift on subsequent use. Furthermore, the same prompt may produce a different output on a repeat query producing inconsistent results. Faculty oversight must therefore be paired with content versioning, periodic rereview, and explicit flagging of AI‐generated material to learners. ²⁸ Free‐tier platforms also lack the data privacy protections necessary for use with patient data in educational settings, and the tools remain unevenly available, with costs, infrastructure requirements, and language biases that limit equitable access.

Trainee overreliance on LLM‐generated study materials, differential diagnoses, or case explanations risks substituting model output for the active reasoning that drives expert development, ²⁹ and the use of generative AI in case write‐ups, board reviews, and competency evaluations raises authorship and academic‐integrity questions that programs should address through explicit policy. As digital pathology becomes integrated into clinical practice, AI literacy will need to become a core competency in cytopathology training for trainees and faculty. Core elements such as case‐curation, question‐writing, and diagnostic‐reasoning skills are themselves teaching assets that should not be progressively delegated to a model.

Research support

The increasing volume and complexity of cytopathology data, spanning digital whole‐slide images, molecular profiling, and multimodal clinical datasets, creates a need for tools that can integrate heterogeneous information. LLMs can assist in this process through literature synthesis, hypothesis generation, and structured data extraction from large report repositories.

For hypothesis generation, LLMs can identify knowledge gaps, suggest testable associations between cytologic features and molecular or clinical variables, and retrieve supporting evidence from published literature. When integrated with pathology foundation models that analyze digital slides, LLMs can add morphologic context to textual and structured data, supporting multimodal investigation. ⁶ , ³⁰ In large‐scale research environments, LLMs can streamline cohort selection by extracting phenotypic patterns from pathology reports and identifying diagnostic trends across institutional databases.

These applications remain largely theoretical in cytopathology. Most demonstrations have occurred in surgical pathology or general medical contexts, and validation on cytology‐specific datasets, including FNA reports, liquid‐based preparations, and organ‐specific classification systems, is needed. Any research use of LLMs also requires attention to the reproducibility of outputs given the stochastic nature of model responses and to the potential for hallucinated citations or fabricated data points; use of LLMs in research is additionally subject to publisher disclosure requirements, and the capacity to generate plausible but fictitious citations, data, and prose constitutes a novel research‐integrity risk that investigators should mitigate through citation verification and provenance logging.

Workflow integration

LLMs can be integrated into laboratory information systems (LIS) to automate data handling and support clinical workflows. LLMs have demonstrated the ability to extract structured data from pathology reports ³¹ ; potential cytopathology‐specific applications include converting narrative cytology reports into structured synoptic formats for LIS entry, extracting specimen adequacy statements, and mapping diagnostic terminology to standardized codes such as SNOMED‐CT and ICD. LLMs have also demonstrated competence in specialized laboratory domains, accurately answering problem‐based queries in clinical chemistry and laboratory management without domain‐specific fine tuning. ³²

In cytopathology workflows, LLMs could identify missing required elements in draft reports (e.g., Bethesda category, adequacy criteria) and prompt users to complete them before sign out. They could also facilitate case triage by flagging reports with a higher probability of discrepancy or clinical significance for prioritized peer review.

Integration with digital pathology platforms and interoperability standards (HL7/FHIR, DICOM‐WSI) would enable LLMs to function within existing clinical infrastructure. Deployment decisions should also weigh the cost of human verification: experience from nonmedical domains shows that LLM‐assisted workflows can net‐decrease productivity when output review takes longer than the task itself, so applications with high task‐to‐verification time ratios (e.g., structured extraction from long narrative reports) are better candidates than those a pathologist could complete directly in less time. ³³ Broader integration challenges in cytopathology AI tools, including data quality, generalizability, and workflow integration, have been recently reviewed. ³⁴

Challenges, limitations, and risks

The potential applications described in Sections 2 and 3 must be weighed against substantial unresolved challenges. Several of these challenges are not merely technical but represent fundamental barriers to safe clinical deployment. This section addresses each in turn.

Hallucinations, accuracy, and patient safety

LLMs generate outputs by predicting statistically likely token sequences, not by reasoning from verified knowledge. This makes them susceptible to hallucination: the production of fluent, confident text that is factually incorrect. In medical contexts, hallucinated outputs can include fabricated citations, incorrect diagnostic criteria, inaccurate drug interactions, or misleading statistical claims. The consequences in cytopathology are direct: an LLM that confidently suggests an incorrect Bethesda category or fabricates a molecular association could lead to inappropriate clinical management.

LLMs are also stochastic when sampling parameters are above zero; the same prompt can produce different outputs on repeated queries, introducing a form of irreproducibility that is incompatible with diagnostic medicine. ¹⁹ Even at temperature zero, outputs remain sensitive to small changes in prompt phrasing, document order, and tokenization, and behavior can shift across model versions, infrastructure providers, and time. To our knowledge, none of the cytopathology‐related studies cited in this review report systematic prompt‐variation or repeated‐query testing, which is a substantive gap. Several strategies can mitigate this nondeterminism. Strategies to produce a more consistent output include: adjusting inference settings to reduce variability; running the same query multiple times and selecting the most common answer, analogous to seeking a second opinion; linking the model to trusted medical databases to anchor responses in established evidence; locking the model version and logging all inputs and outputs for traceability; and treating disagreement across runs as a useful signal of genuine diagnostic uncertainty rather than as noise to be averaged away.

Patient safety concerns extend to patient‐facing applications. General‐purpose LLMs are increasingly used by patients to interpret medical information, with publicly reported adverse events linked to health misinformation generated by these systems. Any suggestion that LLMs can provide patients with “accessible, accurate information” without heavy qualification disregards these documented harms. In thyroid FNAs, for example, the distinction between atypia of undetermined significance and suspicious for malignancy represents a high‐risk diagnostic threshold that can determine whether a patient proceeds to surgical excision; at such a critical juncture, confident‐sounding LLM outputs carry large patient‐safety implications. All clinically related LLM outputs must therefore be subject to rigorous expert review.

No LLMs currently exist that has been trained specifically for cytopathology. General‐purpose models lack the domain knowledge required for reliable performance on specimen‐specific diagnostic tasks, and their outputs require expert verification in all cases.

Explainability and trust

Limited interpretability remains a significant barrier to LLM adoption in pathology. In a global survey of pathologists, 87% reported not using AI algorithms in cytology and 84% reported not using them in surgical pathology, ³⁵ indicating that any novel AI approach, including LLMs, will face cautious adoption. LLMs produce confident outputs without exposing the reasoning that generated them, making it difficult for pathologists to verify whether a conclusion is based on appropriate diagnostic criteria or on spurious statistical patterns.

This opacity is particularly problematic in cytopathology, where diagnoses depend on subtle morphologic features interpreted within clinical context. A pathologist who cannot inspect a model’s reasoning cannot meaningfully integrate its output into a risk‐stratified diagnostic framework.

Several technical approaches address this limitation. Multiturn conversational reasoning, as demonstrated by AMIE, provides stepwise diagnostic explanations that are more transparent than single‐output classifications. ¹⁹ Modular architectures that allow inspection of intermediate processing steps can help pathologists identify where errors originate. Uncertainty quantification and calibrated confidence scores enable models to communicate diagnostic certainty and flag cases where additional information is needed. ⁶ Retrieval‐augmented generation, which grounds LLM outputs in retrieved source documents, can improve factual accuracy and provide verifiable evidence for model responses.

However, explainability in LLMs remains an active research problem, not a solved one. Even models with explanation capabilities may produce plausible‐sounding but incorrect rationales, and clinicians may be susceptible to anchoring bias when presented with confident AI‐generated explanations. ¹⁹ Human‐in‐the‐loop oversight is essential for all current applications.

Bias and generalizability

LLMs trained on large, uncurated datasets risk encoding biases present in their training data. ³⁶ , ³⁷ In cytopathology, this can manifest as better performance on common specimen types (e.g., thyroid FNA) than on less represented categories (e.g., salivary gland or pediatric cytology), and as outputs that reflect the reporting conventions of overrepresented centers and fail on institutions with different staining protocols, scanner hardware, or diagnostic criteria. ³⁸ A model validated at a single academic institution may therefore produce misleading results when deployed in a community or low‐resource setting. ³⁹

Addressing bias requires diverse and representative training data, transparent documentation of data sources and model limitations, and rigorous external validation. Approaches such as domain adaptation, federated learning, and institution‐specific fine‐tuning can improve cross‐site generalizability while preserving data privacy. ⁴⁰ Generalizability itself, however, should not be treated as a default goal in every setting. Tools intended for use within a single institution may benefit from deliberate local fitting. Even aggressive fine‐tuning to the local laboratory case mix, staining protocols, and reporting style may yield better accuracy with more reliable behavior and fewer data‐sharing concerns than a generalized model whose primary benefit accrues to vendors distributing software across sites. Ongoing bias audits across specimen types, patient demographics, and clinical settings should be a condition of deployment regardless of the chosen approach.

Data privacy and security

LLM training and deployment in cytopathology require access to pathology reports, clinical data, and potentially high‐resolution digital slide images, all of which may contain sensitive patient information. Most consumer‐facing commercial LLMs are not HIPAA‐compliant in their default configurations, and free‐tier products in the United States typically retain user inputs for model training and other purposes; HIPAA‐compliant deployment therefore requires an enterprise contract with a Business Associate Agreement, an on‐premises or virtual‐private‐cloud deployment, or use of an institutionally hosted model. Compliance with HIPAA in the United States and GDPR in the European Union establishes the baseline, but pathology data pose additional challenges: the combination of rare diagnoses, high‐resolution whole‐slide images, and molecular data may enable reidentification even from nominally deidentified datasets. ³⁸

Privacy‐preserving training approaches such as federated learning, in which multiple institutions collaboratively improve a model while keeping data at the source, offer an architectural alternative to centralized data pooling. ⁴⁰ , ⁴¹ Cloud‐based model deployment can offer scalability and cost advantages but introduces additional security considerations, including data transit encryption, access controls, and compliance with institutional data governance policies.

Patient autonomy is also a consideration. Even when data are deidentified, patients may reasonably expect transparency about how their specimens contribute to model training, and mechanisms for informed consent or opt‐out (as required under GDPR) support that expectation. ³⁸

Ethical considerations

The deployment of LLMs in cytopathology raises ethical questions that extend beyond data privacy. The principles of beneficence and nonmaleficence require that AI systems demonstrate that their benefits outweigh associated risks before clinical implementation, and the assessment of benefit must extend beyond the patient to the practitioner: a tool that improves a headline metric while degrading pathologist workflow, eroding diagnostic skill, or generating chronic alert fatigue does not satisfy the principle. ³⁸ This standard applies both to patients directly affected by AI‐assisted diagnoses and to those whose data were used in model development.

Accountability is a particular concern. When LLMs support or automate tasks such as structured reporting, differential diagnosis generation, or specimen triage, responsibility for errors becomes ambiguous. The pathologist remains the final decision maker, but reliance on opaque algorithms complicates the chain of clinical reasoning and has implications for both patient safety and legal liability. Institutions deploying LLMs should establish governance structures including oversight committees, ethical review processes, and transparent deployment protocols. ³⁸

Justice and equity considerations require that LLM‐driven tools do not disproportionately benefit well‐resourced institutions while failing to serve underserved populations. If models perform poorly on demographic groups or specimen types underrepresented in training data, their deployment may widen rather than narrow existing healthcare disparities. ⁴² , ⁴³

INFRASTRUCTURE, COST, AND PRACTICAL BARRIERS

LLM deployment carries substantial computational requirements (GPUs/TPUs, energy consumption, on‐premises or cloud infrastructure) that constitute a meaningful barrier for smaller and community‐based laboratories. LIS integration is not standardized across vendors, requiring custom Association for Pathology Informatics and validation layers that must accommodate wide institutional variation in architecture, data formats, and reporting templates. Return on investment in cytopathology has not been established, and cost‐effectiveness data are absent from the literature. Workforce readiness is a parallel concern: effective use requires training in both operation and critical evaluation, without which deployments tend toward either underutilization or uncritical dependence.

Validation gaps

No established framework exists for validating LLMs in cytopathology. Traditional medical device validation assumes deterministic behavior, an assumption LLMs violate through the stochasticity, prompt sensitivity, and version drift discussed previously. Existing pathology validation frameworks provide a useful starting point, but LLM‐specific protocols such as prompt‐variation testing, output‐consistency monitoring, performance thresholds, postdeployment degradation surveillance, and guardrails against harmful outputs remain to be defined. Multi‐institutional prospective validation across diverse specimen types, patient populations, and institutional settings, assessing both accuracy and clinical impact, is needed before any LLM‐based tool is deployed clinically.

Regulatory landscape

No regulations currently target LLM use in cytopathology specifically. However, any medical device incorporating an LLM for cytological analysis falls under existing regulatory frameworks for AI‐enabled software as a medical device. The two principal frameworks, those of the United States Food and Drug Administration (FDA) and the European Union, differ substantially in philosophy and mechanism.

United States: FDA framework

The FDA regulates AI‐enabled software as a medical device through a Total Product Lifecycle approach that treats AI algorithms as evolving rather than static. ⁴⁴ Six AI‐enabled devices specific to pathology had received FDA clearance as of August 2025, though none involves LLM‐based report generation or interpretation support.

The cornerstone of the FDA framework is the Predetermined Change Control Plan (PCCP), which allows manufacturers to obtain preauthorization for a defined set of future algorithm modifications. ⁴⁵ Changes executed according to an approved PCCP and validated within the manufacturer’s Quality Management System do not require a new regulatory submission for each update. The 21st Century Cures Act further clarified that certain software functions are exempt from FDA medical device regulation. ⁴⁶

The FDA has signaled its intent to identify and tag devices incorporating foundation models such as LLMs to improve transparency. ⁴⁷ The prevailing view is that the Total Product Lifecycle and PCCP frameworks are applicable to generative AI in principle, but the burden falls on manufacturers to define verifiable change control protocols for technologies whose behavior is inherently nondeterministic. ⁴⁴

EUROPEAN UNION: AI ACT AND IN VITRO DIAGNOSTIC REGULATION

The EU employs a dual‐layer regulatory structure. AI‐enabled digital pathology tools must comply with both the In Vitro Diagnostic Regulation and the cross‐sectoral AI Act, which entered into force in 2024. ⁴⁸ These legally binding regulations apply uniformly across all member states.

The AI Act introduces a risk‐based classification that designates nearly all AI‐driven diagnostic tools as “high risk.” ⁴⁹ This designation triggers mandatory obligations extending beyond clinical safety to include data governance, cybersecurity, human oversight, and protection of fundamental rights. ⁵⁰

For LLMs, the AI Act operates on two levels. The underlying general‐purpose AI model is subject to baseline transparency obligations, including disclosure of AI‐generated content and publication of training data summaries. ⁵¹ When that model is integrated into a medical device, the entire system becomes a high‐risk AI system subject to the full suite of obligations for risk management, data governance, and human oversight. ⁵⁰ These requirements are particularly challenging for LLMs trained on large, uncurated internet‐derived datasets.

Unlike the FDA’s lifecycle approach that permits iterative updates via PCCP, the EU requires a new conformity assessment by a Notified Body for any substantial modification to a high‐risk AI system. A comparison of both regulatory frameworks can be seen in Table 2. LLMs strain both frameworks. They are resistant to explainability, with internal workings that remain largely opaque even to developers. ⁵² They are also stochastic and susceptible to hallucination, properties that complicate the deterministic validation assumptions underlying current regulatory pathways and leave unresolved how to define, test, and monitor acceptable performance for nondeterministic clinical tools.

TABLE 2.

Comparison of US and EU regulatory frameworks for AI‐enabled medical devices.

Feature	United States (FDA)	European Union
Legal basis	Sector‐specific law via agency guidance 54	Dual‐layer: IVDR + AI Act 51
Philosophy	Lifecycle management, managed flexibility	Precautionary, rights‐based, prescriptive
Risk classification	Device‐based; digital pathology AI is Class II	Dual: IVDR class + automatic “high‐risk” under AI Act
Risk scope	Patient safety and device effectiveness	Patient safety, device performance, and fundamental rights
Algorithm modification	PCCP: preapproved plan for iterative updates	Substantial modification requires new conformity assessment
Oversight body	FDA	National Competent Authorities + Notified Bodies

Role of professional organizations

Professional organizations including the College of American Pathologists (CAP), the American Society of Cytopathology (ASC), the Digital Pathology Association (DPA), the Association for Pathology Informatics (API), and the American Society for Clinical Pathology play a critical role in translating high‐level regulations into laboratory practice. Although these bodies are not regulators, their guidelines and best practices define how laboratories validate and implement new technologies.

The CAP has published guidance on whole‐slide imaging validation and is developing materials on LLM‐specific performance verification, monitoring, and ethical use. The DPA’s regulatory committee engages directly with agencies on digital pathology tool approval.

For manufacturers, regulatory clearance alone is insufficient. Products must also be designed so that clinical laboratories can validate performance and integrate tools into existing workflows according to the standards set by regulatory organizations. ⁵³ This two‐front compliance requirement, regulatory authorization and professional community acceptance, will shape the pathway for LLM adoption in cytopathology.

Recommendations

Based on the current evidence and the challenges outlined in this review, the following priorities for advancing LLM integration in cytopathology are suggested.

• 1) Anchor implementation in patient and practitioner well‐being. The deployment of LLMs in cytopathology should be evaluated against a single overarching question: does this tool improve human health and agency, for patients and for the cytopathologists, cytotechnologists, and laboratory staff who interact with it? Tools that boost a metric while degrading workflow, eroding diagnostic skill, increasing cognitive load, or compromising patient autonomy should not be deployed regardless of headline accuracy. This principle organizes the recommendations that follow.
• 2) Develop cytopathology‐specific benchmarks and validation frameworks. No standardized benchmarks exist for evaluating LLM performance on cytopathology tasks. The field needs curated, multi‐institutional test sets spanning specimen types (FNA, liquid‐based, effusion), classification systems (Bethesda, Paris, Milan), and clinical scenarios. Validation protocols must account for the stochastic nature of LLM outputs, including prompt variation testing and output consistency assessment. Existing pathology AI validation frameworks can be adapted as a starting template for LLM‐specific requirements.
• 3) Prioritize explainability and uncertainty quantification. Clinical adoption depends on trust, and trust requires transparency. LLM‐based tools deployed in cytopathology should provide interpretable rationales, calibrated confidence scores, and clear communication when uncertainty is high. Models should be designed to defer to human judgment rather than present confident outputs indiscriminately. Retrieval‐augmented generation and modular architectures that expose intermediate processing steps represent promising technical directions.
• 4) Build diverse, multi‐institutional training and validation datasets with recognition that local fitting may be preferable for institution‐specific tools. Bias and generalizability cannot be addressed without representative data, and federated learning approaches can support multi‐institutional collaboration while preserving data privacy (Warnat‐Herresthal et al., 2021). For tools intended for use within a single institution, however, deliberate local fitting may yield better accuracy and fewer data‐sharing concerns than a generalized model.
• 5) Begin with low‐risk, high‐value applications. Rather than targeting diagnostic interpretation directly, initial deployments should focus on tasks where LLMs have demonstrated feasibility and where errors are more easily detected: structured data extraction from reports, report completeness checks, terminology standardization, and quality control flagging. These applications provide measurable value while allowing institutions to build experience with LLM tools in controlled settings.
• 6) Establish governance and monitoring infrastructure. Institutions deploying LLMs should create oversight structures that include cytopathologists, informaticists, and patient representatives. These bodies should conduct bias and equity audits across specimen types and demographics, mandate revalidation when data distributions change, and maintain audit trails that can be reviewed during clinical sign‐out. Good machine learning practice principles, including those from joint FDA and Health Canada guidance on data quality, representative sampling, and real‐world monitoring, should be treated as required standards. ⁵³
• 7) Integrate AI literacy into cytopathology training. Both trainees and practicing cytopathologists need structured education on LLM capabilities and limitations: understanding how these models generate outputs, recognizing hallucination and bias, and developing the critical reasoning to evaluate AI‐generated suggestions rather than accept them uncritically. This requirement applies to faculty as well as trainees.
• 8) Engage with regulatory and professional organizations. Cytopathologists should participate in shaping the regulatory frameworks and professional guidelines that will govern LLM deployment. Active engagement with the CAP, ASC, DPA, and regulatory agencies helps ensure that the clinical perspective informs policy development and that emerging standards reflect the practical realities of cytopathology practice.

CONCLUSION

LLMs and their multimodal extensions represent a new category of tools for cytopathology, distinct from the CNN‐based image analysis systems applied to the field in recent years. They offer capabilities in structured reporting, clinical data synthesis, quality control, and interactive query response that address real workflow needs. The evidence base in cytopathology is early, and substantial challenges remain: hallucination risk, limited explainability, bias, data privacy concerns, and the absence of validated frameworks. Responsible adoption will require cytopathology‐specific benchmarks, multi‐institutional validation, transparent governance, and sustained collaboration. With rigorous validation and appropriate oversight, the goal is to augment cytopathology practice for patients and practitioners alike, not to replace expert judgment.

AUTHOR CONTRIBUTIONS

K. Hasan Bilal: Conceptualization; writing — original draft; and writing — review and editing. Joanna Gibson: writing — original draft and writing — review and editing. David Kim: writing — original draft and writing — review and editing. Angelique W. Levi: writing — original draft and writing — review and editing. Aziza Nassar: writing — original draft and writing — review and editing. Handy Oen: writing — original draft and writing — review and editing. Valentina Robila: writing — original draft and writing — review and editing. Mauro T. Saieg: writing — original draft and writing — review and editing. Qiuying Shi: writing — original draft and writing — review and editing. Vanda F. Torous: writing — original draft and writing — review and editing. Akkus Zeynettin: writing — original draft and writing — review and editing. Melissa M. Zhao: writing — original draft and writing — review and editing. Oscar Lin: Conceptualization; writing — original draft; and writing — review and editing.

CONFLICT OF INTEREST STATEMENT

Dr. Oscar Lin is a consultant for Janssen and Hologic. His consultancy work has no impact on the content of this manuscript. All other authors declare no conflicts of interest.