Computational cytology: Bridging microscopy and artificial intelligence

Dear Editor,

Cytology has traditionally relied on the visual interpretation of cellular morphology and the arrangement of cells within a background matrix. For decades, the microscope has been the cytologist’s window to cellular structure, where visual cues, experience, and intuition guided diagnostic decisions are taken.

Today, cytology is evolving into the emerging discipline of computational cytology, where digital imaging merges seamlessly with artificial intelligence (AI). Computational cytology integrates digital image data, patient metadata, and molecular information into machine learning and neural network models.^[1-3] This convergence enables the extraction of complex diagnostic patterns that may lie beyond the limits of human visual perception.

The journey of computational cytology begins with the conversion of traditional glass slides into digital images through whole-slide imaging (WSI). It now provides high-resolution, gigapixel images suitable for analysis through deep learning (DL), one of the most transformative applications of AI in pathology.^[4] Among DL architectures, convolutional neural networks (CNNs) have shown remarkable success in cytopathology, including cervical cancer screening and tumor classification.^[5] Advanced transfer learning (TL) models, such as residual network (ResNet), visual geometry group 16 (VGG16), inceptionv3, and xception, have further improved accuracy and efficiency by leveraging pre-trained networks on large datasets like ImageNet.^[6] The advantages of TL include its ability to work effectively with limited datasets, reduced computational requirements, faster processing time, and robust diagnostic performance. TL-based models have also been applied to predict prognostic markers such as human epidermal growth factor receptor 2 (HER2), Ki-67, and p53, and even to infer molecular alterations like microsatellite instability from routine hematoxylin and eosin-stained histopathology slides.^[7,8] Kather et al. successfully identified microsatellite instability in simple hematoxylin and eosin- stained colorectal carcinoma with the help of ResNet 18 TL.^[7] Coudray et al. predicted mutations of six genes serine/threonine kinase 11 (STK11), epidermal growth factor receptor (EGFR), FAT atypical cadherin 1(FAT1), SET binding protein 1 (SETBP1), kirsten rat sarcoma viral oncogene homolog (KRAS), and tumor protein p53 (TP53) in hematoxylin and eosin-stained histopathology slides of lung carcinomas with the help of inception v3 TL model.^[8] These subtle, sub-visual features, often beyond the perception of the human eye, illustrate how computational cytology expands traditional morphology by linking cytological patterns to molecular and clinical data.

AI as a diagnostic partner

AI should be viewed not as a competitor but as a collaborative partner for cytologists. Computational cytology, enhanced by AI, has the potential to augment diagnostic accuracy and efficiency. AI systems can efficiently screen slides, highlight diagnostically relevant regions, and support tumor classification. DL models can also standardize the scoring of immunocytochemical markers, thereby ensuring greater consistency across observers.^[9-11] Furthermore, natural language processing (NLP) tools can aid in decision support by generating differential diagnoses from descriptive text, cross-referencing relevant literature, and aligning findings with clinical guidelines.^[1] NLP can also extract structured data from narrative reports, capturing tumor types, grades, and prognostic parameters. Importantly, it may help detect discrepancies or inconsistencies within cytology reports and flag potential errors for review, enhancing both quality assurance and patient safety.

From pixels to precision medicine

Computational cytology is advancing rapidly in parallel with developments in genomics and next-generation sequencing. The enormous volume of molecular data generated today exceeds the analytical capacity of human experts alone. Here, DL (particularly CNN-based models) plays an instrumental role in interpreting genomic and transcriptomic information. CNNs can predict variant effects, gene expression patterns, regulatory elements, and exon-intron splice sites.^[1,12-15] The Simplified Molecular Input Line Entry System allows conversion of DNA sequences into one-dimensional images, which CNNs can process through successive convolution, activation, and pooling layers to derive compact feature representations.^[1]

More recently, multimodal data fusion has enabled the integration of diverse data types, imaging, clinical metadata, biomarkers, molecular data, and electronic health records, into unified diagnostic frameworks. This integration supports precision medicine using AI to guide individualized therapy, monitor treatment response, and predict prognosis.^[1]

Future cytology laboratories are likely to evolve into hybrid diagnostic environments, combining WSI scanning stations, molecular pathology platforms, and AI-based analytical systems. Cytologists will collaborate closely with data scientists, bioinformaticians, and clinicians to deliver integrative, patient-centered diagnostics.

CHALLENGES AND ETHICAL CONSIDERATIONS

Despite these advancements, computational cytology faces several practical and ethical challenges. WSI datasets often contain variability in staining, fixation, and image focus, which can reduce algorithmic reliability. DL models are data-intensive, and successful implementation in real-world clinical settings requires large, well-annotated, and diverse datasets.

Data sharing across institutions is essential but raises concerns about privacy and security. Federated learning offers a promising solution, allowing model training on decentralized data where only model parameters, not raw data, are shared with a central server.^[16]

Questions of data ownership and accountability remain unresolved, for instance, whether diagnostic data should belong to the patient, the institution, or the AI developer. The ethical and legal frameworks surrounding data utilization, profit sharing, and intellectual property remain evolving territories.

SUMMARY

Cytology has always balanced art and science. Computational cytology does not disrupt this equilibrium but enriches it. The microscope will remain an indispensable diagnostic tool, while AI provides analytical depth, standardization, and new interpretive possibilities. By integrating imaging, clinical, and molecular data, AI offers cytologists a more holistic view of disease and enhances diagnostic precision and prognostication. The future of cytology lies not only in competition but also in collaboration between cytologists and AI systems. This partnership unites human expertise with computational intelligence to deliver the best outcomes for patients.

AVAILABILITY OF DATA AND MATERIALS

Not applicable.

ABBREVIATIONS

AI: Artificial intelligence

CNN: Convolutional neural network

DL: Deep learning

EGFR: Epidermal growth factor receptor

FAT1: FAT atypical cadherin

Her 2: Human epidermal growth factor receptor 2

KRAS: Kirsten rat sarcoma viral oncogene homolog

NLP: Natural language processing

ResNet: Residual network

STK11: Serine/threonine kinase 11

SETBP1: Set binding protein 1

TL: Transfer learning

TP53: Tumor Protein p53

VGG: Visual geometry group

WSI: Whole-slide imaging

Funding Statement

FUNDING: Not applicable.

AUTHOR CONTRIBUTIONS

PD: Key concept and draft of the manuscript. The author has permitted the release of this version. The author possesses the ability to be responsible for all aspects of the work and ensures that any issues related to the accuracy or completeness of any part of the work are properly investigated and resolved. The author is eligible for ICMJE authorship.

ETHICS APPROVAL AND CONSENT TO PARTICIPATE

Not applicable.

CONFLICTS OF INTEREST

There are no conflicts of interest.

EDITORIAL/PEER REVIEW

To ensure the integrity and highest quality of CytoJournal publications, the review process of this manuscript was conducted under a double-blind model (authors are blinded from reviewers and vice versa) through an automatic online system.