Cytopathology 2.0: How Artificial Intelligence Is Redefining the Future of Cytopathology

Abstract

Artificial intelligence (AI) has driven major disruption across multiple domains of clinical medicine and patient care and is fundamentally redrawing the landscape of modern medicine. In this context, cytopathology stands at a critical crossroads, where traditional microscopic evaluation meets the frontier of computational medicine. Conventionally, a successful cytopathology workflow entails intensive manual effort performed under the close supervision of an expert cytopathologist and an experienced, highly competent team of cytotechnologists. With the advancements in medical science driven by the demand for precision and personalized medicine, workload of the cytopathology laboratory is ever increasing by many folds, while there is an alarming decreasing trend in the availability of skilled human resource. A new era of diagnostic precision is emerging, as machine learning and deep learning algorithms take center stage in the laboratory. These systems tend to streamline the workflow by reviewing high-volume slide sets, prioritizing high-risk cases, and offering prognostic insights. These processes are highly dependent on meticulous digitization of cytology smears using whole slide imaging pathology scanners, which in turn enables telecytology, large-scale data sharing, and the development of robust training datasets, all of which accelerate AI innovation. This review provides an overview of the technical processes and applications of AI-based cytopathology algorithms across different organ systems, workflow transformation (from preanalytical to quality control, telecytopathology, and integration with molecular diagnostics), and the various challenges and limitations in their adoption in the routine diagnostic workflow for patient care.

INTRODUCTION

Cytopathology has long been a cornerstone of clinical medicine, especially in its utility in the rapid diagnostic process of cancer screening and early detection of infectious diseases. The microscopic examination of cells under the light microscope is often subjective and variable and highly dependent on the experience and expertise of the observer, since it requires an extraordinary degree of pattern recognition and cognitive endurance. Globally, human factors have become a major impediment, considering the increase in cytopathology volumes and scarcity of trained pathologists.

With the emergence of artificial intelligence (AI) as a major transformative force across medical disciplines, pathology in general, and cytopathology in particular, stands at the crossroads of this digital revolution. By integrating AI, particularly machine learning (ML), deep learning (DL), and generative AI (Gen AI) algorithms, there is a scope for a paradigm shift in the current workflow. Thus, computational models capable of recognizing complex morphological patterns can be leveraged, and these AI models can analyze enormous datasets without training, triage suspicious cases, and even predict clinical outcomes.[1,2,3,4,5]

Digital pathology (DP) is crucial for this transformative process, since digitized glass slides of cytology specimens enable telecytopathology, large-scale data sharing, and the development of robust training datasets, all of which in turn accelerate the initiative of AI innovation. The present-day advanced neural networks integrate with high-resolution whole slide imaging (WSI) systems to detect subtle cellular and nuclear features of malignancy, selectively identify and sort abnormal cell populations in a smear, and optimize laboratory workflows. These increased capabilities have had a significant influence on DP, and there is now a clear forward-looking perspective in which foundation models are taking control. There are hybrid models consisting of vision transformers (ViTs) and convolutional neural networks (CNNs), which have enormous potential in establishing cytopathology as a crucial part of precision oncology.[6,7,8,9,10,11,12,13]

This evolution from manual workflow and microscopy to AI-enabled diagnostics assures not only enhanced diagnostic accuracy and sensitivity but also addresses the global shortage of specialized cytopathologists by streamlining the routine screening process. Further, in the era of “multi-cytomics,” cytomorphological data can be seamlessly integrated with clinical and molecular data to form an integral part of patient triage and management.[5,6,7,8,9,10]

TECHNICAL FOUNDATIONS OF AI IN CYTOPATHOLOGY

AI is best understood as a broad field of computer science that focuses on creating systems that are able to perform tasks that typically require human intelligence. These applications include, but are not limited to, visual perception, speech recognition, decision-making, and language translation. The evolution of AI in cytopathology is best perceived as a hierarchical progression of technologies that progresses from general pattern recognition to autonomous, goal-oriented reasoning.

At this juncture, it is essential to understand various terminologies such as ML, DL, Gen AI, and Agentic AI, which are used to describe the structural hierarchy, as well as artificial neural network (ANN), CNN, and ViT, which are used in the context of neural network architectures.[3,4,5,6,7,8,9,10,11]

The Structural Hierarchy: From AI to DL

To understand how these terminologies relate within the structural hierarchy, they can be considered as a set of nesting dolls: AI is the largest category, ML is a subset of AI, DL is a subset of ML, and Gen AI is a specific application of DL [Figure 1]. In effect, at its broadest level, AI encompasses any computational system capable of simulating human cognitive functions, within which ML represents a shift toward algorithms. These algorithms tend to develop through data exposure rather than rigid, manual programming. DL, a specialized subset of ML, utilizes ANNs, which are computational architectures inspired by the biological neural networks of the human brain, to process high-dimensional complex data, such as whole slide images that the human brain cannot process.

Figure 1.

Diagrammatic representation of the hierarchical relationship between artificial intelligence (AI), machine learning, deep learning, Generative AI, and Agentic AI

To further elaborate on each of the components of the structural hierarchy [Table 1]:

Table 1.

Summary of AI models in cytopathology

Technology	Core mechanism	Clinical application in cytology
CNN	Local filter scanning	Detection of malignant cells and nuclear features
ViT	Global self-attention	Assessment of cellular arrangement and architectural patterns
Gen AI	Data synthesis	Creation of synthetic training sets for rare pathologies
Agentic AI	Autonomous reasoning	Integrated workflow management and multistep diagnostic assistance

AI = artificial intelligence, CNN = convolutional neural network, Gen AI = generative artificial intelligence, ViT = vision transformer

• AI: This is the principal concept of machines mimicking human cognitive functions. It can be broadly classified into the following two categories:

  • Narrow AI (Weak AI): This is the currently used form of AI, which is designed to perform a specific task (e.g., facial recognition or identification of a specific cell type in a pathology slide).

  • General AI (Strong AI): This is a hypothetical AI that possesses the ability to understand, learn, and apply knowledge across any domain, while functioning at a human level.

• ML: This subset of AI focuses on using data and algorithms to imitate the way humans learn and tends to gradually improve its functional accuracy without being explicitly programmed for every scenario. This system identifies patterns in data to make predictions, rather than operating within the strict confines of “if-then” rules. So, ML is capable of predicting the probability of a patient’s response to a specific treatment based on historical clinical data.

• DL: This is a specialized subset of ML based on ANN with multiple layers (hence termed “deep”). This system tends to mimic the structure of the human brain to process data in a nonlinear way. It is particularly powerful for unstructured data, such as images and text, and is capable of analyzing a high-resolution digital slide to segment tumor cells from healthy tissue.

• Gen AI: This subtype of AI is trained on vast amounts of existing data to learn the underlying patterns and is thus able to generate “new” examples that resemble the original data, such as text, images, audio, or synthetic data. A suitable example is that of large language models or tools that generate synthetic cytological images for training purposes when real cases are rare. In research, Gen AI is increasingly used to supplement rare datasets (e.g., rare malignancies) to provide the necessary volume of training data that traditional collections may lack.

• Agentic AI: This is one of the most significant shifts in modern AI development. Unlike traditional AI, which waits for a prompt and provides a single answer, Agentic AI acts as an “agent” capable of independent reasoning and multistep execution. It can function autonomously to break down a complex goal into smaller tasks; choose to use external tools (e.g., searching a database, running a Python script, or browsing the web); and also perform self-correction to review its own work and adjust its strategy if it encounters an error. In the cytopathology workflow, an Agentic AI system would not only identify suspicious cells in a smear but would also independently retrieve the patient’s prior genomic or clinical history and draft a comprehensive diagnostic report for pathologist review.

Neural Network Architectures: ANN, CNN, and ViT

Though the term “ANN” encompasses all these layered architectures, three distinct models have defined the progress of digital cytology:

• Standard ANN (Multilayer Perceptron): This is the foundational model where nodes are fully connected and is functionally formidable for structured data (e.g., predicting patient risk based on demographics or biomarkers). However, it lacks the spatial awareness required for complex image analysis.

• CNN: Currently, this is the “gold standard” for medical imaging. This model utilizes convolutional filters to scan images for diagnostic features such as nuclear contours, chromatin patterns, and nucleoli. They are exceptionally robust at identifying specific cellular morphology within a “grid-like” pixel structure and have therefore produced desired outcomes in various aspects such as cell localization, region separation, and classification in liquid-based cytology (LBC) samples.

• ViT: This is a more recent revolutionary development, which breaks slides into discrete patches and then analyzes them simultaneously. By virtue of applying “Self-Attention” mechanisms to images, ViTs excel at capturing the global context and long-range spatial relationships between cell clusters, which is vital for assessing the cytoarchitecture in cytology smears. ViT is capable of addressing the limitations of CNN, which focus on small local areas of the image without having the ability to capture tissue patterns and the spatial orientation of cells that hinder their potential in diagnosing a broader picture. On the contrary, ViTs are developed particularly to capture a full global representation of an image by splitting it into patches and understanding how cells have interactions with distant cells, as well as complexities, such as overlapping cell patterns.[6] These have proven to be a more viable option than traditional methods, as they can capture the entire image at once and provide a holistic understanding of complex cytological patterns. ViTs are highly valuable in identifying cytological heterogeneity, as in cytological evaluation of thyroid aspirates.

As we observe new developments in AI, there is an increasing demand for multimodel systems that combine the local feature identification capability of CNNs with the relationship-based analysis of ViTs. By combining CNNs and ViTs, these hybrid architectures successfully capture both fine-grained local details and global contextual relationships, resulting in highly accurate analyses.[8]

Foundation models have been discovered to be trained on a broad and complicated data collection utilizing a self-supervised technique. They are not required to learn from every manually annotated dataset. They adapt and learn from a huge pool of observations, eliminating the need for extensive training on well-labeled data. In the domain of cytopathology, differences may be observed in how laboratories and hospitals carry out their routine activities. Because foundation models are effective at learning from a huge pool of datasets rather than only from the annotated data, they have broader knowledge, which helps to mitigate issues in data-limited environments.[5]

Phases of Development of an AI Algorithm in Digital Cytopathology Workflow

The development of an AI algorithm in cytopathology can be broadly divided into three phases, which are briefly summarized below and diagrammatically represented in Figure 2:

Figure 2.

Schematic representation of artificial intelligence (AI) in digital cytopathology spanning slide scanning, data curation, labeling and annotation, and model optimization and validation

• Phase A (Data Acquisition): The workflow begins with the high-throughput digital scanning of multimodal physical slides into a central repository. These large-scale images are then “patched,” that is, broken down into smaller, manageable digital segments and zoomed to high magnification to capture fine cellular details.

• Phase B (Annotation and Curation): Experts perform “AI-assisted labeling,” where human pathologists and AI copilots work in tandem to establish a diagnostic consensus. This high-quality, verified data are then organized into a structured digital biobank, categorized by diagnostic difficulty (e.g., borderline or rare cases).

• Phase C (Optimization and Validation): The curated data are processed through complex DL architectures. Through data augmentation and federated learning protocols, the model is trained and tested. The final output is an optimized detection system, evaluated via a comprehensive dashboard that tracks performance metrics such as F1 scores and area under the curve (AUC).

ORGAN-SPECIFIC APPLICATIONS OF AI IN CYTOPATHOLOGY

Gynecological cytology: Cervical smears

Cervical cancer is one of the most common malignancies affecting women worldwide.[14] Screening by cervical smears remains one of the most common strategies for the early detection and prevention of cervical cancers. The sheer volume of screening and the manual effort involved have made cervical Pap smear examination one of the earliest and most extensively studied modalities in the context of AI and automation.

The automated Papanicolaou test has been in existence for over 30 years, with PAPNET (1992) being the first commercially available automated screening system approved as a method of rescreening slides marked as negative by cytologists.[15] This was followed by a Food and Drug Administration (FDA)-approved ThinPrep imaging system in 2004, which detected and segregated smears with abnormal cells for manual screening by cytotechnologists.[16] Numerous other automation systems, such as the FocalPoint GS imaging system, followed suit, but the cost-effectiveness for implementing at large-scale population screening became a crucial limiting factor in its adoption.[17,18]

Recently, with the rapid development in the domain of AI, DL models such as CNNs have been widely applied to conventional Pap smears as well as LBC smears.[19,20]

Segmentation of cervical smears

There are five major aspects of automation in slide screening: image acquisition, preprocessing, segmentation, feature extraction, and classification of cervical cytology cells.[21] The prerequisite for screening using AI is focused on nuclear and cytoplasmic abnormalities for segmentation.[21,22,23] Researchers in this area used supervised learning and focused on extracting adaptive shape from cytoplasmic contour fragments along with shape statistics for the segmentation of overlapped cytoplasm in cervical smear images.[24] A range of methods have been used for the study of AI in cervical cell segmentation, ranging from the mean-shift clustering algorithm proposed by Wang et al.[25] to the patch-based CNN models. The other algorithms included CNN-based models proposed by Gautam et al.[22] and Song et al.,[24] as well as the superpixel-based Markov random field model,[26] all of which demonstrated that the subjective shortcomings of the manual segmentation process could be successfully addressed by AI and automation to detect abnormal cells with precision.[21]

Classification of cervical cancer cells

Another crucial step in cervical cancer screening is the accurate classification of cervical cancer cells in smears. This is generally met with the limitation of a considerable degree of interobserver variability, depending on the expertise and experience of the observer.[27] This poses a serious challenge in regions where a trained cytopathologist may not be available, thus creating an opportunity for AI. Thus, many classification systems have been proposed in the past to tackle this issue. These have been majorly dependent on morphometric characteristics of the nucleus and the cytoplasm, along with the texture features of the cells and their background.[25,28,29] Most studies observed a significant reduction in the time for evaluation, along with an increased efficiency and limited observer bias.[21]

Other groups focused on the DL framework and graph convolution network instead of the traditional segmentation characteristics for classification, and showed high diagnostic accuracy.[30,31] While most of the studies are retrospective in nature, a few prospective trials have demonstrated consistent improvements in detecting high-risk lesions while reducing the time for screening.[32,33]

Performance in detection and classification

Overall, across all studies, AI-based screening was observed to surpass manual methods, with workflows significantly streamlined and assessment and triaging time reduced [Table 2].[19,20,32,34,35,36,37,38]

Table 2.

Comparison of performance of AI models in cervical smear screening across key studies

Study	AI model	Dataset	Key metrics (sensitivity/specificity/AUC)	Lesion detected
[19]	AICCS/deep learning	16,056 slides	94.6%/89.0%/0.947	NILM to HSIL cytology grades
[20]	AICyte/CNN	32,451 slides	99.3%/9.87% standalone	HSIL/SCC
[35]	CNN for LBC	1,605 whole slide images	NA/NA/0.89–0.96	Neoplastic vs. non-neoplastic
[36]	AttFPN/attention CNN	7,030 slides	95.83%/94.81%/0.991	Abnormal cells
[37]	Deep learner classifier (p16/Ki-67)	High; reduced need for colposcopies	NA	Precancer in dual stain
[38]	CNN	2,816 LBC images	95.63%/79.85%/NA	Identification of abnormal foci from LBCC smears

AI = artificial intelligence, AICCS = artificial intelligence cervical cancer screening, AICyte = artificial Intelligence cytology system, AttFPN = attention feature pyramid network, AUC = area under the curve, CNN = convolutional neural network, HSIL = high-grade squamous intraepithelial lesion, LBC = liquid-based cytology, LBCC = liquid-based cervical cytology, NA = not applicable, NILM = negative for intraepithelial lesion or malignancy, SCC = squamous cell carcinoma

Though AI shows promise for potential point-of-care screening, it is recommended that a broader, multiethnic dataset study needs to be conducted across diverse population groups to circumvent the study biases arising from single-institution-based research.[33,35]

Thyroid cytology

Thyroid cytology constitutes a major workload in most cytopathology laboratories, owing to which a lot of AI-based studies have been undertaken. DL models have shown a strong potential in enhancing the diagnostic accuracy of thyroid cytopathology. Numerous retrospective and prospective studies have been conducted on thyroid fine needle aspiration (FNA) smears with the purpose of automating the classification of benign versus malignant nodules and achieving a direction in indeterminate Bethesda categories [Table 3].[39,40,41,42,43,44,45,46,47,48]

Table 3.

Comparison of AI models in their application in thyroid cytopathology

Study	AI model	Dataset	Key metrics (sensitivity/specificity/AUC)	Lesion detected
[43]	CNN + ANN hybrid	Meta-analysis (studies from 2000 to 2023)	91%–98%/NA/0.92–1.00	Indeterminate nodule triage
[39]	Inception ResNet v2 (CNN)	10,332 patches/306 FNA smears	100%/90.4%/0.97	Classification of FNA smears
[41]	ANN	87 cases	90.48% sensitivity; 83.33% accuracy	Papillary carcinoma detection
[44]	Backpropagation ANN	57 cases	Successfully distinguished all cases of FA from FC	Distinguishing FC from FA based on nuclear and cytological morphometric features
[45]	ANN	355 cases	Correctly distinguished between benign and malignant lesions	Benign vs. malignant lesions
[46]	ANN and nuclear morphometry	197 cases of thyroid follicular tumor	Successfully identified FA from FC in 87% of the cases using ANN and 97% of the cases using morphometry	FC vs. FA
[47]	ANN with nuclear morphometry	157 cases		Benign vs. malignant lesions
[48]	Feed-forward neural network with cytological features and clinical data as input nodes	453 cases	ANN had higher sensitivity and a lower specificity when compared to cytology	Benign vs. Malignant

AI = artificial intelligence, ANN = artificial neural network, AUC = area under the curve, CNN = convolutional neural network, FA = follicular adenoma, FC = follicular carcinoma, FNA = fine needle aspiration, NA = not applicable

One of the studies by Lee et al.[39] provided remarkable evidence for a CNN-based system in differentiating papillary thyroid carcinomas from benign nodules, where AI reduced the interobserver variability, with kappa increasing from 0.64 to 0.84 post-assistance.

Few systematic reviews have been conducted to observe the use of AI in the resolution of the gray areas of thyroid cytopathology. Poursina et al.[43] observed an improved categorization and risk stratification using AI-assisted assessment of quantifiable cytological markers such as nuclear grooving and nuclear inclusions in the indeterminate cases. A similar result was observed by Girolami et al.,[42] where their systematic review highlighted the role of AI in resolving atypia of undetermined significance through WSI by addressing the subjective biases of manual evaluation. A better triaging of the smears has thus been observed to reduce unnecessary surgeries for benign nodules.[41]

Few studies used ANN models to distinguish between thyroid follicular tumors, achieving over 90% accuracy in the process.[44,46] Furthermore, emerging multimodal AI systems showing an integration of cytology with molecular studies of BRAF/NRAS mutations have shown promising results in achieving high sensitivity by needle rinse genotyping without the need for cell-block preparation.[49]

AI models have displayed considerable potential to standardize Bethesda reporting and reduce the rates of reporting of “indeterminate” category cases significantly.[42] However, broader validation and standardization are required to realize the full impact of adopting AI-based algorithms in the reporting of thyroid cytology for routine patient care.

Breast cytology

AI in breast cytopathology has evolved from early ANN models to advanced DL frameworks. Early work employed ANN models trained on nuclear morphology features extracted from FNA samples, achieving 100% sensitivity and specificity in differentiating fibroadenoma from invasive ductal carcinoma (IDC).[50,51]

One of the earliest studies using ANN and Bayesian analysis was conducted by Dawson et al.,[52] who differentiated high-grade from low-grade breast carcinomas using image analysis-based nuclear grading of the cytological preparations. In the same decade, Einstein et al.[53] used both logistic regression analysis and an ANN based on the fractal geometry of breast carcinoma. The development of early-stage neural networks is based on image analysis and automated learning systems, such as the Xcyt developed by Teague et al.[54] Further, reviews on computer-aided diagnosis (CAD) also formed a bedrock in the understanding and development of various ML algorithms [Table 4].[51,55,56,57,58,59]

Table 4.

Comparison of studies using AI models in breast cytopathology

Study	AI model	Dataset	Key metric	Lesion
[56]	ANN	64 FNAC lesions		ILC vs. IDC vs. benign
[57]	Visiopharm AI algorithm	105 cytology specimens with metastatic breast carcinoma	Concordance between pathologists’ and AI scores: 94.3%	Automated ER AI algorithm to differentiate metastatic breast carcinoma
[58]	ANN (Neurointelligence software)	472 cases		Benign vs. malignant lesions
[59]	BFCNet (CNN)	1,020 region-of-interest patches from Giemsa-stained slides 631 region-of-interest patches from H&E-stained slides	Accuracy: Giemsa-stained: 97.53% H&E-stained: 96.59%	Diagnosis of ductal carcinoma

AI = artificial intelligence, ANN = artificial neural network, BFCNet = Breast FNAC Classification Network, CNN = convolutional neural network, ER = estrogen receptor, FNAC = fine needle aspiration cytology, H&E = hematoxylin and eosin, IDC = infiltrating ductal carcinoma, ILC = infiltrating lobular carcinoma

Following this, in 2013, Dey et al.[56] applied ANN to identify lobular carcinoma, where 64 smears of histopathologically proven breast lesions were selected and subjected to automated image morphometry. The model was successful in identifying benign IDC and the majority of invasive lobular carcinoma. This set a precedent for ANN to be developed in future for an unknown set of qualitative and quantitative data.[56] Further, a few authors used ANN based on cytomorphological data, morphometric data, nuclear densitometric data, and gray-level co-occurrence matrix to measure image texture and assist in the diagnosis of gray-zone breast lesions of FNA cytology (FNAC).[58,60]

DL architectures, especially CNNs, have been applied to WSIs of breast cytology to automate the detection of atypical cells and provide risk stratification. These models primarily identify subtle nuclear pleomorphism, chromatin patterns, and architectural abnormalities, which often escape routine visual assessment. Studies have reported accuracies up to 90% in binary classification.[61] These advances have enabled a better understanding of CNN-based models, leading to the development of models such as the Breast FNAC Classification Network (BFCNet) developed by Bal et al.[59] However, the primary limitations of such models remain the limited size and the lack of external validation.[61]

Further, the Visiopharm AI algorithm has been successfully validated for immunohistochemistry (IHC) analysis of estrogen receptor (ER) expression in whole slide images of metastatic breast carcinoma, with an excellent concordance between the pathologists’ and AI scores (94.3%).[57]

Respiratory system cytology

Respiratory cytopathology encompasses a vast spectrum of samples, including sputum analysis, bronchial washings, bronchial brushings, endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA), and guided FNA. AI is being increasingly applied in this area; however, its use and study remain nascent compared with gynecological cytology [Table 5].[62,63,64,65,66,67,68,69,70]

Table 5.

Comparison of AI models in key studies in respiratory cytopathology

Study	AI model	Dataset	Key metrics	Lesion studied
[64]	DCNN: DenseNet121	37 cases	Malignancy prediction in patch-level classification: accuracy: 0.94, sensitivity: 0.97, and specificity: 0.85 SqCC prediction in patch-level classification: accuracy: 0.90, sensitivity: 0.86, and specificity: 0.91	Prediction of malignancy and histological type of lung cancer
[65]	Six models: VGG19, ResNet50, ResNext50, InceptionV3, DenseNet121, and EfficientNetB7	Multicenter study, 1,273 whole slide images	Increased pathologist accuracy to 96.9%	Benign vs. malignant
[66]	Fine-tuned deep CNN (AlexNet, GoogLeNet, VGG16, and ResNet50)	55 cases	Classification accuracy: • AlexNet: 74% • GoogLeNet: 66.8% • VGG16: 76.8% • ResNet50: 74%	Classification of cytological diagnosis of lung cancer
[67]	CNN – ResNet101	97 patients	95.5% accuracy in classification 98.8% sensitivity and specificity	ROSE in EBUS for identifying malignant cells
[68]	CNN based on Google Inception V3	40 cases	AUC = 1 (Diff-Quick, Pap-stained slides) AUC = 0.875 (H&E slide) 100% sensitivity	SCLC vs. LCNEC
[69]	VGG16 model DCNN	46 cases	Accuracy: 87%	Benign vs. malignant

AI = artificial intelligence, AUC = area under the curve, CNN = convolutional neural network, DCNN = deep convolutional neural network, EBUS = endobronchial ultrasound, H&E = hematoxylin and eosin, LCNEC = large cell neuroendocrine carcinoma, ROSE = Rapid On-Site Evaluation, SCLC = small cell lung carcinoma, SqCC = squamous cell carcinoma

Despite the need for cytopathologists in precision medicine, an inadequate distribution of professionals leads to most of the cases presenting at advanced stages, as they evade early diagnosis.[70] Thus, automation in this field is needed to address this lacuna of early detection and reduce the time of diagnosis, thereby facilitating early institution of definitive management, leading to a reduction in morbidity and mortality.[69]

A major retrospective multicenter study conducted in Korea by Kim et al.[65] in 2023 evaluated six representative CNN models (VGG19, ResNet50, ResNext50, InceptionV3, DenseNet121, and EfficientNetB7). The study focused on interobserver variability in the diagnosis of lung cytopathology, along with the effect of AI as a potential augmentation tool to balance out the human observer bias.[65] They observed the benefit of AI as a second-opinion tool for trainees, which may reduce the cost and time to educate cytopathologists. However, the AI model was observed to overdiagnose artifacts such as mucin material and pyknotic cells against a bloody background mimicking dyskeratotic squamous cells. Some false-negative cases misdiagnosed by AI were identified by pathologists, thus showing the black box-like nature of the current AI models.[65]

Another area under active investigation in thoracic cytopathology with immense potential for AI integration is EBUS-TBNA. DL models are being applied to EBUS images for automated lymph node detection, localization, and malignancy risk assessment for optimal targeting of biopsies.[71]

AI-assisted Rapid On-Site Evaluation (ROSE) during EBUS-TBNA holds particular promise for improving the adequacy of specimens and reducing nondiagnostic rates by providing immediate feedback on the cellular content in the preliminary morphological assessment. This would be especially beneficial in resource-limited settings with a lack of on-site trained cytopathologists.[67,71]

CNN-based models are being evaluated for subtyping of primary lung cancers.[61] The study by Gonzalez et al.[68] was one of the first to use a DL algorithm to resolve the diagnostic dilemma of distinguishing large cell neuroendocrine carcinoma from small cell lung carcinoma (SCLC) on cytology smears. Tsukamoto et al.[66] used four fine-tuned deep CNNs (DCNNs: AlexNet, GoogLeNet, VGG16, and ResNet50) for automated classification of lung cancer using images.

Another study by Tanaka et al.[64] used a DCNN on LBC samples to classify adenocarcinomas from squamous cell carcinomas. In this, the Dense-Net121 architecture was employed, which used multiple layers (convolutional, pooling, and fully connected) to automatically learn and extract complex features from images. These models were designed to identify subtle morphological patterns that might be difficult for human eyes to categorize consistently, often achieving diagnostic accuracy comparable to or exceeding that of experienced cytopathologists in differentiating adenocarcinomas from squamous cell carcinomas and normal cells. The patch-level accuracy to differentiate malignant versus normal was 94% specificity, and the case-level accuracy was 91%; while corresponding figures for adenocarcinoma versus squamous cell carcinoma were 90% and 78%, respectively, indicating a scope for further improvement in future studies.

Recently, studies combining cytology with ancillary techniques and DP have been increasingly reported. Combining DP with multiplex immunofluorescence (IF)/IHC assays for PD-L1 assessment in non-SCLC has been proposed for use in cytological samples.[72] Ishii et al.[72] also predicted EGFR, KRAS, and ALK alterations, reporting patch-level accuracy above 90%.

Major hurdles posed are the preanalytical variables, such as inconsistent staining of sputum samples, which reduce the generalizability of AI across centers.[62] Further, limited respiratory-specific labeled datasets may potentially lead to biases in groups that are underrepresented, such as nonsmokers or early-stage lesions.[63] Another pitfall is the lack of alignment between AI models and the Papanicolaou Society recommendations for respiratory cytology, where the current protocols undervalue the role of AI in ancillary testing.[73,74]

Pancreaticobiliary cytology

Pancreaticobiliary tumors are one of the most aggressive cancers and have a poor outcome owing to their late stage of diagnosis. To address this concern, AI has emerged as a promising tool in the diagnosis and characterization of dysplastic and neoplastic lesions, along with cyst fluid analysis.[75] The rapid pace of development of predictive models holds the potential to increase the diagnostic accuracy while providing a fair probability analysis of the benefits of therapeutic interventions.[75]

To resolve the diagnostic challenges posed by pancreatic lesions owing to their anatomic location, morphology, and unique function, few authors have attempted to use AI models in their studies. One of the earliest studies was conducted by Momeni-Boroujeni et al.,[76] who used a K-means clustering algorithm with a multilayer perceptron neural network (MNN) model to segment the cell clusters into separable regions of interest, followed by the extraction of the areas for cytological diagnosis in 75 FNA biopsies to differentiate pancreatic lesions as benign or malignant. The accuracy for the benign and malignant categories was 100%, while that of the atypical dataset was 77%. Their study thus highlighted the potential of using the MNN model for image analysis.[76]

Another notable study by Lin et al.[77] demonstrated the efficacy of AI as a substitute for manual ROSE during endoscopic ultrasound-guided FNA. They addressed the existing lacunae due to the unavailability of ROSE in Asian and European institutions, owing to a shortage of trained manpower.[77] A similar study was conducted by Zhang et al.,[78] where 5,345 cytopathological slide images were subjected to analysis by a DCNN system to segment cell clusters and identify tumor cells. The AUC was observed to be >0.900 in the sensitivity analysis and was observed to be superior to that of trained endoscopists, comparable to that of their cytopathologists.

Kurita et al.[79] compared the diagnostic ability of carcinoembryonic antigen (CEA), cytology, and AI in diagnosing cystic lesions from pancreatic cyst fluid. A retrospective analysis of 85 patients was conducted with a DL diagnostic algorithm constructed to account for a wide variety of factors, such as serum CEA, CA19.9, CA125, and amylase levels, along with the anatomic and cytological details of the cyst. The accuracy of the AI model was observed to be higher than that of CEA and cytology. Thus, they observed that AI may improve the ability to diagnose pancreatic cystic lesions accurately [Table 6].[79]

Table 6.

Comparison of AI models in key studies of pancreatic cytopathology

Study	AI model	Dataset	Key metric	Lesion studied
[76]	MNN	75 FNA cases	100% accuracy for malignant 77% accuracy in an atypical dataset	Benign vs. malignant cases
[77]	ResNet101V2 neural network	693 images	88.7% accuracy 78% sensitivity 90.6% specificity	AI model for ROSE during EUS-FNA
[78]	DCNN system (U-Net-based CNN with ResNet101 encoder/decoder)	5,345 cytopathological slide images	AUC > 0.900	AI model for ROSE during EUS-FNA
[79]	ANN	85 pancreatic cyst fluid samples	AUC in CEA: 0.719 AUC in AI: 0.966 Sensitivity of AI: 95.7% Specificity of AI: 91.9% Accuracy of AI: 92.9%	Diagnostic ability of CEA cytology and AI in differentiating benign vs. malignant cyst lesions

AI = artificial intelligence, ANN = artificial neural network, AUC = area under the curve, CEA = carcinoembryonic antigen, CNN = convolutional neural network, DCNN = deep convolutional neural network, EUS-FNA = endoscopic ultrasound-guided fine needle aspiration, MNN = multilayer perceptron neural network, ROSE = Rapid On-Site Evaluation

Exfoliative cytology

Urinary system

Urothelial carcinomas (UCs), one of the most common cancers worldwide, are generally multifocal with a tendency to recur post-treatment.[14] With this in context, urine cytology, being an effective, noninvasive, and inexpensive modality, takes the center stage in the screening and surveillance of UCs.[80] Thus, the integration of AI models targeting cell classification and risk stratification, while aligning with the Paris System for Reporting Urinary Cytology, has become the focus of most studies.[81]

One of the studies that stood out was that of Nojima et al.,[82] who developed a novel DL system (DLS) for detecting high-grade UC (HGUC) cells using gradient-weighted class activation mapping, which highlighted the nuclear color tones as a key diagnostic indicator for HGUC. This system used a pretrained 16-layer Visual Geometry Group (VGG16) CNN to analyze Papanicolaou-stained urine cytology images and achieved excellent diagnostic performance for detecting HGUC cells, with an AUC of 0.9890, an accuracy of 95.62%, and an F1 score of 0.9071. Since the annotation of malignant cells in the study was performed by a pathologist, it was deemed that the DLS diagnoses were within the correct limits.[82] Another major merit of the study was the ability of their DLS to not only accurately detect UC cells but also to distinguish histological characteristics such as stromal invasion on cytological screening. This is generally not feasible by classical cytology due to the absence of stroma in the specimens. The evidence of stromal invasion forms the basis of classifying the cancers as superficial or invasive, further affecting therapeutic decisions. In this regard, the authors regarded DLS to be superior to classical cytology.[82] Another study predicted muscle invasion of the tumor using attention-based multiple-instance learning on digitized slides.[81]

Similarly, other groups have designed studies using DLS and have observed a strong cytology–histopathology correlation.[81,83,84,85] CNN networks were used to extract certain marked features from the cytology images, followed by identification and demarcation between atypical urothelial cells and various background elements such as inflammatory cells.[85]

Although AI algorithms demonstrated good sensitivity for detecting neoplastic cells, specificity lagged in some studies due to artifact-induced false positives.[83] Further, interobserver agreement was observed to improve along with a reduction in screening time.[83] However, a pitfall in accuracy was noted by Muralidaran et al.,[86] where one of the low-grade UC smears was misclassified as high-grade [Table 7].

Table 7.

Comparison of AI models in their application in urine cytopathology

Study	AI model	Dataset	Key metrics (sensitivity/specificity/AUC)	Lesion detected
[81]	ResNet50 + attention-based multiple instance learning	105 slides	89%/62%/0.83	HGUC vs. non-HGUC
[85]	Deep learning on WSI	786 whole slide images of liquid-based cytology	NA/NA/0.984–0.990	Neoplastic vs. non-neoplastic
[82]	VGG16 CNN	466 images	90.0%/100%/0.989 (binary) 82.4%/72.9%/0.866 (high grade)	Malignant vs. benign; invasive/high-grade prediction
[86]	ANN	115 urine cytology samples	—	Malignant vs. benign
[87]	ANN	85 cases	Successfully distinguished between benign and malignant cases	Benign vs. low-grade UCC vs. high-grade UCC

AI = artificial intelligence, ANN = artificial neural network, AUC = area under the curve, CNN = convolutional neural network, HGUC = high-grade urothelial carcinoma, UCC = urothelial carcinoma, WSI = whole slide imaging

Further, the emerging use of AI is also being studied in predicting the invasiveness of the tumor based on neutrophil count. However, it is cautioned to note that the potential conflicts due to artifacts need to be taken into account while interpreting the results.[88]

A few AI software systems such as VisioCyt, AlxURO, and CellsVision are currently in commercial use for evaluating urinary cytology for HGUC and have been reported to offer high diagnostic performance.[89] Overall, these models hold promise for precision-based diagnosis of urine cytology smears, potentially reducing cystoscopies and invasive biopsies for more refined patient care to a larger population base and at a modest cost.

Effusion cytology

Effusion cytology involves the examination of fluid samples from body cavities, generally to identify malignant cells. However, overlapping features between reactive mesothelial cells and neoplastic cells are major confounders and potential challenges to the cytopathologist, with diagnostic errors in up to 20%–30% of cases, without any ancillary tools.[90] In a routine cytopathology laboratory, these cases undergo extensive workup with immunocytochemistry or IHC of cell blocks, which not only burdens resource-constrained setups but also increases the time to diagnosis and cost of diagnosis.

Thus, recently, there has been an increasing trend to automate the detection and classification of malignant cells in pleural and peritoneal fluid samples for optimal workup. Currently, the DL models applied to effusion cytology smear workflow have shown an accuracy of over 85% in distinguishing adenocarcinomas from benign or reactive mesothelial cells.[91,92] Limited work was done using an ANN on effusion cytology. Truong et al.[93] used an ANN model based on densitometric and morphometric data for the diagnosis of lymphocyte-rich effusions, showing a sensitivity of 95% and specificity of 85.7%. Another study by Barwad et al.[94] used an ANN model using cytological features and image morphometric data as input to identify malignant cases.

The current AI models primarily focus on image-based analysis of cytological smears and their cell blocks for identifying features of malignancy, such as nuclear atypia and clustering. CNN variants such as U-Net for semantic segmentation and YOLOv8 for object identification have been used for delineation of cells and their subsequent classification into adenocarcinoma cells.[92,95] Reset-50 and AlexNet models have been used to classify neoplastic versus non-neoplastic cells, achieving an accuracy of 88%–89% using morphometric, colorimetric, and texture characteristics.[96]

Most of the evidence in effusion cytology is limited to pleural effusion, with few studies on peritoneal or mixed fluids.[96] This constitutes algorithmic biases from imbalanced datasets, which can be solved by diverse multicentric population studies. Further, a lacuna was observed in the use of real-time AI deployment in routine laboratories or long-term outcomes on patient survival, which can be overcome should more cytology laboratories opt to adopt an AI-based workflow after extensive validation.

Hematolymphoid

Hematology

DL models such as CNNs have been widely applied to automate hematological cell detection and classification, improving diagnostic accuracy for hematological disorders such as acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS).[97,98] The whole slide images of bone marrow (BM) aspirates are analyzed to identify the individual cells, generate histograms of different cell types, and reduce interobserver variability.[98]

A DL analysis model termed the Cell Detection and Confirmation Network was proposed by Su et al.[99] to aid in the diagnosis of AML. Another study proposes a cell separation algorithm, achieving an average accuracy of 95% in the segmentation process, with the extracted nuclear and cytoplasmic data obtained in this process used to classify leukemia and its subtypes.[100] An accuracy of 84% was achieved in lymphoblastic subtypes and 92% in myeloblastic subtypes. Differences in the color and texture characteristics formed the basis of this differentiation.[100] Another CNN-based model by Eckhardt et al.[101] characterized AML in BM aspirates and predicted NPM1 mutational status with an accuracy of 0.86, using similar prediction pattern associated with condensed chromatin and perinuclear lightening dysplasia pattern.

Further, AI-assisted models such as CellaVision and Morphogo are used in hematology to examine peripheral blood smears and BM aspirates, respectively.[50] The CellaVision application uses ANN to identify blasts with a sensitivity of 100% and specificity of 94%, as well as white blood cells and common red blood cell (RBC) changes, including target cells, with a high degree of accuracy.[102] The drawback, however, lies in its inability to differentiate between myeloblasts, lymphoblasts, and monoblasts.[50]

Morphogo is a 27-layer CNN AI system trained on a vast dataset of over 2.8 million BM nucleated cell images and is used for BM aspirate analysis, which differentiates diverse nucleated cells in high-resolution images.[103] This model was used by Fu et al.,[104] who reported an accuracy of 85.7%–91% in the classification and analysis of 230 cases of hematopoietic lineage cells. This system was also used for automated megakaryocyte identification in BM aspirates with a high sensitivity (96.5%) and specificity (89.71%).[105] A retrospective review of 60 cases also used the Morphogo software to evaluate and detect non-hematopoietic malignancies in BM aspirates with an accuracy of 82.2%, sensitivity of 56.6%, and specificity of 91.3%.[106]

Another comprehensive review studying the use of CNN designs, hybrid models, and ensemble techniques in improving acute lymphoblastic leukemia diagnosis and classification observed the merits of the specialized CNN architectures. However, the study was limited by the small size of the datasets, and the authors proposed the incorporation of molecular and genomic data in future analysis.[107]

While leukemia cells have nuclear and cytoplasmic details enabling easier differentiation, RBCs, being anucleate, pose a different challenge. Kaji et al.[108] developed an AI system to diagnose cytogenetically defined MDS from BM smears, while solving this issue. They addressed segmentation challenges, such as overlapping cells, by binary mask analysis of RBCs, thus revealing the shape and distribution patterns associated with MDS. These findings propose AI assessment of RBC morphology as a novel biomarker for MDS.[108]

Thus, while AI models reduce variability, subtle morphological overlaps in the hematological cells pose a challenge and call for larger prospective validation studies.[109] Further, most of the studies were focused on hematological malignancies. A diverse dataset, including non-leukemic disorders are hence needed to prevent biases.[109] This would enable a more holistic incorporation and streamlining of digital systems and remote analysis while integrating AI with clinical workflows, as has been attempted in the study by Bermejo-Palaez et al.[110]

Lymph node

Owing to its simplicity and early time of diagnosis, FNAC is a key stone in the diagnostic algorithm in lymphadenopathies. To make the process more efficient and effective, DCNN models such as Inceptionv3 have been studied for use in lymph node cytopathology.[111] Another study by Beriwal et al.[112] used a three-classification model, that is CNN, InceptionNet, and ResNet50, to detect lymph node malignancies in whole slide images after due preprocessing of the images. They noted the ResNet50 as the most accurate model out of the three.[112]

However, as noted by both authors, clinical practice leads to an encounter with a diverse spectrum of challenging cases, which highlight the need for larger studies with a broad dataset to compare the performance of different AI models.[111,112] This is further emphasized by the paucity of studies using AI in lymph node cytopathology.

Miscellaneous

Application of AI remains relatively underdeveloped in certain organ systems such as soft tissue, salivary gland, and the lymphoreticular system.

The morphological overlap between the various soft tissue neoplasms poses an inherent diagnostic complexity, which is further amplified by the rarity of many tumor subtypes. For definitive classification of these lesions, ancillary studies such as IHC and molecular testing are critical and pose a significant challenge for algorithm training.

Similarly, a review of literature showed that while studies on the application of AI in the histopathology of salivary glands are being undertaken, studies on AI models in salivary gland cytopathology are still in their emerging stage.[113]

DP AND WORKFLOW TRANSFORMATION

The shift to digital cytology

With the introduction of WSI scanners in the cytopathology laboratories, there has been a gradual shift in the workflow of traditional cytology. Now, smears (either conventional cytology or LBC) on glass slides are scanned and digitized. It is important to note that preanalytical workflows are crucial steps in DP, where automated slide labeling and barcoding are essential before scanning of slides.[114] Scanned whole slide images are uploaded to a centralized Image Management System (IMS),[114] which primarily helps in remote access, data exchange, and telecytology-based consultations.[115] When the IMS is linked with the laboratory information systems (LIS) and the electronic health record, cytopathologists can access patient data, review the slide, and provide a diagnostic report from anywhere in the globe without having to wait for the slides to be shipped, which saves time and expenses, and increases efficiency.[116] AI algorithms that have undergone a stringent validation protocol can be applied to the digitized images to screen and flag suspicious cells for final review by cytopathologists, thus enhancing screening accuracy, reproducibility, and efficiency and reducing work-related fatigue. As a result, it has been observed that utilizing AI reduces observer variability and has a strong potential for early-stage detection.

Telecytopathology and access to expertise

Telepathology is a subset of telecytopathology in which we use cytology samples, such as FNA smears, to do real-time analyses. It is performed during the process of aspiration and involves staining and microscopic scanning of smears, allowing the cytopathologist to offer real-time feedback. The ultimate purpose of completing a real-time ROSE is to ensure that the sample aspirated is sufficient for diagnosis without having to repeat the process.[115] With recent improvements, ROSE may now be carried out in remote or network hospitals, virtually by sending digital images to cytopathologists for a quick clinical interpretation. AI algorithms are being developed to assist in screening of ROSE smears for a more efficient workflow process.

Quality assurance, education, and training

It is imperative to understand that the success of any DP or telepathology system, and developing AI algorithms on images acquired on these systems, warrants a robust quality assurance process. This is a critical prerequisite in DP and AI algorithm-based workflows. The framework is based on defined operating processes, system navigation, workflow integration, and ongoing performance monitoring. The cytopathology laboratory team should undergo regular training programs for skill upgradation and awareness about developments in the domain. The training should include knowledge of the digital workflows, validation protocols, image handling steps, and troubleshooting workflow challenges, as they emerge. Working across many cross-functional teams and offering feedback jointly can help reduce errors. Regular audits should be performed to ensure adherence to quality standards and regulatory guidelines.[114]

Integration with ancillary and molecular testing

Cytology is developing, and its purpose is growing beyond morphological assessment to include comprehensive molecular testing. Cell blocks prepared from aspirates can be used in ancillary (IHC) and molecular testing methods such as polymerase chain reaction, fluorescence in situ hybridization assays, and next-generation sequencing (NGS) for accurate characterization and biomarker prediction: for example, ER, progesterone receptor, and Her2 status in breast carcinoma; BRAF and RAS in thyroid nodules; and EGFR and ALK alterations in lung cancer.[117,118] Molecular testing can help identify risk factors and guide personalized treatment recommendations. The combination of morphological evaluation and molecular data has paved the way for precision medicine in the field of cytology. AI algorithms can help integrate clinical information with cytomorphological, histopathological, and molecular genetic data for a holistic workup of the patient.

Challenges and limitations

Despite considerable advances and promise, the adoption of AI in cytopathology is fraught with multiple hurdles that distinguish it from other fields of DP. While the potential for increased efficiency is high, the transition from glass slides to digital images has multiple challenges and limitations:

1. Related to scanning of cytology smears: Unlike histopathology slides, where tissue sections are relatively flat, cytopathology involves three-dimensional cellular clusters.

   • The “depth-of-field” problem: Conventional cytology smears often contain thick groups of cells. Standard WSI may capture only a single plane, leading to out-of-focus areas that render AI analysis impossible.

   • Z-stacking demands: To overcome this, scanners must use “Z-stacking” (capturing multiple focal planes).[13] However, the acquisition of multiple images causes a significant increase in file sizes and scanning time, straining information technology infrastructure and subsequent storage budgets. This facility is available with only a few scanners and is often a major impediment in the workflow of cytopathology laboratories.

   • Artifacts and variability: Smear thickness, staining variations across different laboratories, and the presence of background “noise” (such as blood, inflammation, or mucus) can confuse AI algorithms that are not sufficiently robust.

2. Related to interpretation and validation: One of the most significant barriers to clinical trust is the lack of transparency in DL.

   • Lack of justification of diagnosis generated: Many AI models provide a classification (e.g., “Malignant”) without explaining why they reached that conclusion. In a clinical setting, a cytopathologist signing off the report must be able to verify the morphological features that generated the result by the AI algorithm.

   • Validation issues: Algorithms trained on one set of data may perform poorly when introduced to slides from a different hospital with different staining protocols, a phenomenon known as “domain shift.”

3. Related to scarcity of datasets and annotation: AI requires massive amounts of high-quality, labeled data to learn effectively.

   • Lack of adequate number of cytopathologists for annotation: It requires a large team of expert cytopathologists to manually annotate thousands of individual cells or clusters on a digital screen, a process that is significantly more time-consuming and exhausting than traditional microscopy.

   • Rare entities: Developing AI for rare malignancies is difficult because there are simply not enough digital cases available to train the model adequately.

4. Related to regulatory and ethical concerns: The path from a research laboratory to a clinical workspace is heavily regulated.[119,120,121,122,123]

   • Liability: If an AI model misses a high-grade lesion (a false negative), the question of legal liability—whether it lies with the pathologist, the hospital, or the software developer—remains largely unsettled.

   • FDA and global approval: Navigating the regulatory requirements for “Software as a Medical Device” (SaMD) is a costly and lengthy process that can slow down the deployment of innovations.

5. Related to workflow integration and cost: The “Digital Pathology Gap” is often financial and logistical.

   • High upfront costs: Implementing AI requires a massive investment in high-speed scanners, high-resolution monitors, and powerful servers, as well as high recurrent costs of storage. Many laboratories find it difficult to justify these costs without a clear return on investment.

   • Disruption of routine: Integrating AI into a high-volume laboratory requires a total overhaul of the traditional workflow. Any technical lag or “downtime” in the digital system can cause significant delays in patient reporting.

Regulatory and accreditation frameworks for AI in cytopathology

Recognition of AI algorithms by regulatory bodies and accreditation agencies is progressing, but unevenly across jurisdictions and specimen types.

• In the United States, Current Procedural Terminology codes 0827T–0856T, introduced in 2024, now cover cytopathology services performed alongside AI algorithms.[119] AI tools are regulated by the FDA as SaMD, and the number of cleared tools in DP has expanded significantly as of May 2026.

  • In histopathology, SaMD tools that have now been approved include Paige Prostate Detect (De Novo, 2021), Ibex Prostate Detect, formerly Galen Second Read (510(k), 2025), ArteraAI Prostate (De Novo, 2025), ArteraAI Breast, cleared in May 2026 as the first FDA-cleared DP-based risk stratification tool for early-stage HR-positive/HER2-negative breast cancer,[120] and PathAI’s AISight Dx, a DP IMS cleared via 510(k) in June 2025 for primary diagnosis with an authorized Predetermined Change Control Plan.[121]

  • In cytopathology specifically, the Hologic Genius Digital Diagnostics System, with the Genius Cervical AI algorithm, received FDA clearance in 2024 as the first and only FDA-cleared digital cytology system, combining DL AI with advanced volumetric imaging to identify precancerous lesions and cervical cancer cells from ThinPrep Pap tests.[122] This system functions as an AI-assisted review tool, not as an autonomous primary screener. Despite the growth in histopathology, cytopathology-specific clearances beyond cervical screening remain absent, underscoring a persistent validation dataset gap in this subspecialty.[123]

• In India, the National Accreditation Board for Testing and Calibration Laboratories–accredited laboratories follow ISO 15189:2022, requiring documented method validation and LIS audit trails for AI outputs, with concordance study thresholds set locally.

• In the United Kingdom, RCPath and the National Pathology Imaging Co-operative are building national imaging datasets for AI benchmarking. However, cytopathology-specific AI mandates remain under development, pending the National Commission on the Regulation of AI in Healthcare’s recommendations, which is expected in 2026.

All three frameworks share a common expectation: AI must be validated prospectively within each laboratory’s own workflow. Performance must be monitored over time across different stains, preparation methods, and scanners. A recurring problem is that many AI models are trained on isolated, well-selected cell images rather than representative clinical slides, limiting their real-world generalizability.[124] The EU AI Act (Regulation (EU) 2024/1689) has reinforced this by classifying AI diagnostic tools as high-risk medical devices.[124] Concordance studies must therefore cover a broad range of specimen types, with a focus on diagnostically critical categories such as malignant and atypical cells.

FUTURE DIRECTIONS

With the advent of Agentic AI, the future of AI-assisted cytology is poised for greater transformation, including performing virtual staining (routine, histochemistry, and IHC), generating virtual cytology images, executing virtual molecular profiling, and developing AI-assisted and AI-collaborative workflows.

Virtual staining

Currently, multiple groups are doing focused work to introduce AI-assisted slide staining in cytopathology.[125,126] Although most of the effort has been for histopathology, the potential of extrapolating this technology to cytopathology smears and cell blocks is immense. This process of “virtual staining” utilizes advanced generative models such as Generative Adversarial Networks to digitally transform unlabeled or traditionally stained tissue images into various target modalities. Available AI algorithms, such as “RestainNet,” employ self-supervised learning to act as digital re-stainers, converting grayscale or hematoxylin and eosin (H&E)-stained images into high-fidelity representations that preserve cellular structure while correcting for color inconsistency. These help in normalizing analytical bias due to variation in staining protocols adopted by various laboratories, so that scanned digital images are consistent while performing validation steps or subsequent analysis.

Similarly, platforms such as “DeepLIIF” (Deep Learning Inferred ImmunoFluorescence) are virtual restaining and quantification AI models for converting standard IHC images into more informative multiplex IF formats. This has been shown to enhance the dynamic range and sensitivity for biomarkers such as “HER2-low” in breast carcinoma cases. In the commercial sector, Pictor Labs has developed products including ClearStain, DeepStain, and ReStain. The latter enables the virtual transformation of H&E-stained images into histochemically stained images, such as those stained with Masson’s trichrome or periodic acid–Schiff, without using serial sections of precious tissue. The AI platform of ViewsML converts routine H&E slides into high-resolution biomarker data in just minutes, bypassing physical IHC and the need to consume valuable tissue, expensive reagents, and intense workflow. These models are of immense value when the material is scarce and tissue needs to be preserved for subsequent molecular sequencing. These computational approaches not only streamline laboratory workflows by reducing reagent dependency but also provide a critical preprocessing step for CAD systems by standardizing image appearance across different scanner types and staining protocols.

Generation of virtual cytology images

Though cytopathology is the cornerstone of tissue diagnosis, progress is hampered by the limited data availability and privacy regulations. To circumvent these challenges, Zheng et al.[127] have developed COIN, a controllable cytology image generation foundation model that provides a robust and privacy-preserving framework for generating scalable cytology data. This model can synthesize realistic images, thus providing a valuable tool to accelerate the development and implementation of AI-based diagnostic solutions.[127]

Virtual molecular profiling

With the advancement of computational inference, AI has been integrated into virtual RNA and DNA workflows, transforming molecular and cytopathology. These workflows leverage “Image-to-Molecular” models that predict genetic and transcriptomic profiles directly from standard H&E-stained digital slides.

Virtual transcriptomics (RNA-Seq prediction)

AI models, such as HE2RNA, are trained to predict whole transcriptome profiles from whole slide images by performing “Spatial Mapping,” where AI analyzes images tile-by-tile and generates virtual spatial transcriptomics heatmaps. This reveals how gene expression varies across different regions of a tumor, without the requirement of expensive spatial sequencing. Effectively, this promises to offer a rapid and economical way to reveal molecular blueprints of immunotherapy biomarkers (such as PD-L1) or microsatellite instability directly from routine pathology slides and also to assess tumor heterogeneity. Leading-edge frameworks such as Spatial Expression-Aligned Learning now allow for the systematic alignment of local morphology with spatially resolved transcriptomic data, enabling “virtual spatial transcriptomics” that can predict localized gene expression profiles from image patches alone. This morphomolecular coupling is particularly vital in cytopathology, where samples such as FNAs are often limited in volume.[128,129,130,131]

Virtual genomics (DNA mutation prediction)

AI-powered histology is now a standard tool for inferring specific molecular alterations without consuming tissue for NGS. “Genotype–phenotype mapping” models can predict critical mutations, such as IDH status in gliomas, EGFR in lung cancer, or BRAF in melanoma, with high accuracy based on cellular architecture and nuclear atypia. This is very useful in cases with limited biopsy material (as in cytology cell blocks), where AI identifies which slides are most likely to yield positive molecular results, ensuring efficient tissue stewardship.[132,133]

AI-assisted cytology workflow

Currently, an AI-driven workflow requires human intervention to define the task and its execution. Unlike this workflow, Agentic AI systems have an innate ability to understand, plan, and execute complex tasks autonomously in a functional clinical diagnostic cytopathology laboratory. These intelligent agents would be capable of planning, managing, and executing the entire workflow, from slide scanning to preliminary analysis and generating comprehensive diagnostic reports, while highlighting suspicious areas for final review by an expert. DL models such as FNA-Net can now perform in situ adequacy screening on unstained slides, predicting if a sample will have sufficient DNA/RNA for downstream molecular testing before it even reaches the lab.

AI-collaborative workflow

This would be a paradigm shift in AI-assisted workflow to AI-collaborative workflow, where Agentic AI plays the role of an autonomous partner, thereby significantly enhancing the efficiency, speed, and accuracy and unraveling diagnostic insights which are currently beyond human capacity. Advanced systems (e.g., Tempus xT) fuse imaging, genomic, and transcriptomic data into a single multimodal integration framework, empowering pathologists to perform “multimodal reasoning,” thereby linking morphological impression directly to its genetic risk stratification, and also helping to predict tissue viability, forecast drug efficacies using 3D tumor organoids, and identify clinically actionable gene fusions. Furthermore, Agentic AI could facilitate continuous learning and adaptation, integrate new research findings, and evolve diagnostic criteria in real time, ensuring that cytopathology remains at the forefront of medical diagnostics.

CONCLUSION

The integration of AI into cytopathology represents a transformative shift toward a more objective, efficient, and data-driven diagnostic model. AI is currently redefining the boundaries of cytopathology and surging toward a future of enhanced accuracy, reproducibility, and efficiency. Beyond routine diagnostics, AI-driven algorithms have the potential to expand the scope of conventional cytopathology into precision healthcare by incorporating predictive analytics, personalized medicine, and integrating with multi-omics data.

Success will depend on navigating significant critical hurdles, including standardizing datasets, mitigating algorithmic bias, rigorous validation across diverse populations, and the establishment of clear regulatory and ethical frameworks. Apprehension among pathologists and the need for upgrading technical knowledge constitute a major factor in its ready adoption across multiple cytopathology laboratories. Practicing cytopathologists should be educated and reassured that AI is not intended to replace them; rather, it serves as a powerful auxiliary tool that allows the human expert to focus on complex diagnostic dilemmas, and AI algorithms manage the high-volume routine screening work. The future of cytopathology lies in this symbiotic relationship, where technology handles the scale, and the pathologist provides the definitive judgment. However, as DL models continue to mature, the synergy between human expertise and computational precision will undoubtedly redefine the future of cellular diagnostics and precision medicine, ultimately advancing patient care through more reliable and personalized diagnostic pathways.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.