Truace
Subscribe

CYTOLONE-v2 hierarchical AI classification for urinary and cervical cytology

Source article: A WSI-Independent Inference Framework for Cervical and Urinary Cytology Using Hierarchical Labels

Background Cytology artificial intelligence (AI) systems often depend on whole-slide images (WSIs) or high-magnification cell images and are commonly optimized for a single specimen type. We developed CYTOLONE-v2, a contrastive language-image pretraining (CLIP)-based hierarchical classification framework for cervical and urinary cytology that is designed for individual-field inference without WSI input. Methods CYTOLONE-v1.1, previously trained for cervical cytology, was further trained by low-rank adaptation us…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2074· cite & verify
A WSI-Independent Inference Framework for Cervical and Urinary Cytology Using Hierarchical Labels
202306 Cervical Cancer Screening by DataBase Center for Life Science (DBCLS). CC BY 4.0 · https://creativecommons.org/licenses/by/4.0
Trace impact readingNegative state
P 71The P score combines the specificity and measured human impact of the grounded problem claim with the strength of this Trace’s cited sources.

Both sides are scored from claims and sources, not community votes.

G 65The G score combines the specificity and measured human impact of the grounded gain claim with the strength of this Trace’s cited sources.

In brief

Researchers built CYTOLONE-v2, a CLIP-based hierarchical classifier meant to work on individual microscope fields rather than whole-slide images. They took a cervical cytology model and further trained it with low-rank adaptation on urinary cytology fields at roughly 10x magnification, then tested it on 40 independent urinary cases.

The work shows that one framework can plausibly span two specimen types, which matters because many cytology AI tools are tied to a single specimen and to whole-slide scanning. Its urinary specificity was low and cervical performance dropped at finer levels, so it is not yet shown to be useful in clinical practice.

Main points

  1. Urinary field-level accuracy was 64.0% for Anomaly and Malignancy and 72.5% for Diagnosis.
  2. Sensitivity for urothelial carcinoma was 99.0%, but specificity against reactive change was 46.0%.
  3. Cervical accuracy at the Anomaly level was retained (94.8% vs. 94.7%), and Malignancy-level accuracy fell from 93.5% to 91.2%.
  4. The 90-case source cohort was split into 72 training and 18 validation cases, with five diagnosis-representative fields chosen manually per case.

The gain

Extending a cervical cytology model to urinary cytology gave 99.0% sensitivity for urothelial carcinoma at the Diagnosis level in single-field inference without whole-slide input.

The problem

The same urinary model had only 46.0% specificity against reactive change, and cervical accuracy fell at the Bethesda and Diagnosis levels after adaptation.

The rundown

The framework starts from CYTOLONE-v1.1, which had been trained for cervical cytology. It was adapted using WSI-derived urinary fields equivalent to 10x magnification, and the split was done at the case level to separate training and validation cases.

The urinary test cohort held 20 reactive change cases and 20 urothelial carcinoma cases, giving 200 fields. The high sensitivity with low specificity means the model flagged carcinoma well but often misclassified reactive change.

On the independent cervical test set, the model kept its Anomaly-level accuracy and lost a little at the Malignancy level. The drop was larger at the Bethesda and Diagnosis levels, so extending the model to a new specimen type cost some performance on the original one.

What this doesn’t fix

Findings are limited to selected WSI-derived fields and a reactive change versus urothelial carcinoma test cohort; external, case-level, direct-microscopy and prospective validation is still needed.

Sources

  1. Peer-reviewedCytopathology2026-10-20

The debate