Cervical Cytology Screening

The highest-volume, most protocolised, most automatable task in pathology — and the one place where AI is already deployed at national scale, which makes it the best available evidence on what deployed pathology AI actually looks like.

The highest-volume, most protocolised, most automatable task in pathology — and the one place where AI is already deployed at national scale, which makes it the best available evidence on what deployed pathology AI actually looks like.

What it is

Cervical screening reads a liquid-based or conventional smear for abnormal squamous and glandular cells and reports it against a fixed vocabulary — the Bethesda System, which gives a small closed set of categories (NILM, ASC-US, LSIL, ASC-H, HSIL, AGC, and so on). Volumes are enormous, the vast majority of slides are negative, and the task is screening rather than diagnosis: the cost of a missed HSIL is very different from the cost of an unnecessary referral.

Those four properties together — high volume, low prevalence, closed vocabulary, asymmetric costs — are why this is the most automated task in the discipline, and why it is worth a page here even though the group does not currently do it.

Why it is a different technical problem from histology AI

The unit of evidence is a cell, not a region. A smear is a dispersed population; a single abnormal cell among tens of thousands decides the report. That is closer to a needle-in-a-haystack detection problem than to the tissue-architecture problems most of this wiki is about, and it is why the field’s methods cluster around cell detection and per-cell classification rather than Weakly Supervised Learning over tiles.

Slide-level inference has to aggregate from a very small number of positive instances. Several of the works below are explicitly about the patch-to-slide step under extreme imbalance, which is the same structural problem as MIL but at a far worse positive rate. See Class Imbalance and Accuracy for why an accuracy figure on a screening cohort says almost nothing.

Scanning is not the same operation. A smear is thick and cells overlap, so a single focal plane does not capture everything a screener would examine by racking focus — cytology WSI conventionally needs multiple focal planes (a z-stack), which multiplies scan time and storage per slide. Whether the department’s scanner can produce z-stacked cytology at acceptable throughput is a prerequisite question for any of this and is not recorded anywhere here. [unverified]

Entry points

Pointers taken from Awesome-AI4DigitalPathology, not appraisals. None has been read here and nothing on this page claims a result from any of them; they are recorded so that the cost of entering the topic is a decision rather than a search.

Reviews first, since this is a mature field:

  • Computational Cytology Survey, Med Image Anal. Paper
  • Cervical Cytology Deep Learning Review, Neurocomputing 2024. Paper

Datasets, which are unusually good in this area and mostly open:

  • SIPaKMeD — Pap smear cells for classification, on Kaggle. Paper · Data
  • HiCervix — hierarchical cervical cytology benchmark. Paper · Code
  • HMCHH-TCT-CellDet — a large annotated ThinPrep dataset, Sci Data 2025. Paper · Data
  • BMT — cross-validated ThinPrep Pap dataset, Sci Data 2024. Paper

Deployed or near-deployed systems, which are the reason this page exists:

  • AIATBS — AI-assisted Bethesda classification, Nat Commun 2021. Paper
  • Cervical WSI Screening — Nat Commun 2021, with code. Paper · Code
  • AICCS — screening system, Nat Commun 2024. Paper · Code
  • DualCytoNet — explicitly targeted at low-resource settings, Nat Commun 2025. Paper
  • LBC-DL — liquid-based cytology for precancer detection, Nat Commun 2025. Paper · Code
  • Smart-CCS — pretraining plus test-time adaptation. Paper · Code
  • UniCAS — a foundation model specifically for cervical cytology, Cell Rep Med 2025. Paper · Code

That five of these are in Nature Communications between 2021 and 2025, several with released code, is itself the finding: this is a field with an unusual density of large, published, code-released clinical systems.

Why it matters for my work

It is the best available answer to “what does deployed pathology AI actually look like?” Every AI project in this wiki — PinkKidney, Liver Tru-Cut — Tumour Origin AI, Paige Prostate, Aiforia Breast — is at the internal-validation stage. Cervical screening is the one application where systems have been built, published with code, and put into service at scale. If the question is what the path from model to deployment looks like, this is where the worked examples are, and they are free to read.

The economics are the argument, and they are legible. Turnaround Time and Laboratory Workload Measurement both concern making a case for capacity from measurement. Screening is where that case is easiest to make, because the workload is high, repetitive and dominated by negatives — which is exactly the shape where triage has value. Whether the department does cervical screening at volume is not recorded here and decides whether any of this is more than background.

It is a clean instance of an argument this wiki makes repeatedly. Reader variability in cytology is well known and sets the ceiling on any reference standard — the Interobserver Agreement argument, in a setting where the categories are fixed and the disagreement is therefore measurable rather than diffuse.

And it is where the group’s existing instincts transfer directly. The nucleus-segmentation canon that Cell Detection and Counting is about — HoVer-Net, Cellpose, CellViT, PanNuke, MoNuSeg, CoNIC — sits in this same section of the source list, because cell-level detection is the shared substrate. Anyone who has done Ki-67 counting has done a version of this task.

How it connects

Cell Detection and Counting — the shared substrate: per-cell detection and classification is what both this and Ki-67-style indices reduce to.

Class Imbalance and Accuracy — a screening cohort is the canonical imbalanced population, and the page that explains why an accuracy figure here is close to meaningless.

Interobserver Agreement — the ceiling on any cytology reference standard, in the setting where it is most studied.

Whole Slide Imaging — cytology scanning differs from histology scanning in a way that changes scan time and storage; the z-stack requirement is the specific issue.

Turnaround Time and Laboratory Workload Measurement — the operational case for automating a high-volume screening workload, which is where this becomes a departmental question rather than a research one.

Ki-67 Proliferation Index — the other counting task in this wiki, and the closest analogue in technique if not in purpose.

External Validation — the deployed systems above are the wiki’s best examples of models that did face outside cohorts, which is what most projects here have not.

Open questions

  • Does the department do cervical cytology at volume? Nothing in this wiki records it, and it decides whether this page is a project or background reading. [unverified]
  • Can the available scanner produce z-stacked cytology slides at usable throughput, and what does that do to scan time and storage? Scanner Logs and Scanning Time in Real Life hold the machinery to answer the throughput half from data already collected.
  • Is there a Turkish or institutional Bethesda-coded archive that could serve as a cohort? The public datasets above are ThinPrep-heavy and mostly Chinese in origin, so case mix and preparation method may not transfer.
  • Of the deployed systems, which report an external cohort rather than an internal split? Not checked — and it is the question that would make one of them a model to follow rather than another internal-validation result. [unverified]