kuri54/mlx-CYTOLONE-v2

CYTOLONE-v2

📖 Overview

CYTOLONE-v2 extends CYTOLONE-v1.1 from cervical cytology to cervical and urinary cytology using a shared CLIP-based model and specimen-specific hierarchical labels.

The model was developed in the study “A WSI-Independent Inference Framework for Cervical and Urinary Cytology Using Hierarchical Labels.”

CYTOLONE-v2 classifies individual low-magnification cytology fields. It is designed to accept individual images without requiring a whole-slide image (WSI) at inference. However, all development and evaluation images in the study were extracted from WSIs; direct microscope-camera performance has not yet been validated.

This is a research model for field-level cytomorphologic classification, not an autonomous screening system or a substitute for professional cytological diagnosis.

📝 Model Description

• Architecture: CLIP ViT-B/16, originally derived from openai/clip-vit-base-patch16.
• Starting checkpoint: CYTOLONE-v1.1, previously trained for cervical cytology.
• Adaptation: Additional training with urinary cytology data using low-rank adaptation (LoRA) in both the vision and text encoders.
• Distributed weights: A full checkpoint with the LoRA adapters merged into the model. Separate adapter loading is not required.
• Input resolution: 224 × 224 pixels after standard CLIP preprocessing.
• Output: Image–text similarity scores for the selected candidate labels. Softmax scores depend on the candidate set and should not be interpreted as calibrated probabilities of cancer.

The repository retains the mlx-CYTOLONE naming convention. The CYTOLONE v2 application loads this checkpoint with Transformers/PyTorch and uses MPS acceleration on supported Macs.

🧬 Dataset and Label Structure

📊 Urinary Dataset Composition

The urinary dataset comprised Papanicolaou-stained liquid-based cytology specimens collected at three institutions. Five diagnosis-representative fields were manually selected per case.

Diagnosis	Training cases	Validation cases	Independent test cases
Negative	24	6	0
Reactive change	24	6	20
Urothelial carcinoma	24	6	20
Total cases	72	18	40
Selected fields	360	90	200

The 90-case source cohort was split into training and validation subsets at the case level (80:20; seed 64). All five fields from each case remained in the same subset. The independent test cohort contained no cases shared with the source cohort.

Fields were extracted at 1024 × 1024 pixels and 0.92 µm/pixel, corresponding to approximately a ×10 objective field. Standard CLIP resizing and normalization were applied. No color normalization or data augmentation was used.

The reference standard was the routine final cytological diagnosis at each institution, not uniform histological confirmation. Fields were selected with knowledge of that diagnosis, before model inference and without access to model outputs. The evaluation therefore concerns selected-field classification, not systematic slide screening.

🏷️ Hierarchical Labels

Cervical cytology uses four levels: Anomaly → Malignancy → System/Bethesda → Diagnosis.

Normal Benign NILM Negative
Normal Benign NILM Atrophy
Anomaly Dysplasia LSIL Mild_dysplasia
Anomaly Dysplasia HSIL Moderate_dysplasia
Anomaly Dysplasia HSIL Severe_dysplasia
Anomaly Carcinoma SCC Squamous_cell_carcinoma
Anomaly Carcinoma ADC Adeno_carcinoma

Urinary cytology uses three levels: Anomaly → Malignancy → Diagnosis.

Normal Benign Negative
Normal Benign Reactive_change
Anomaly Malignant Urothelial_carcinoma

For urinary cytology, Normal denotes the computational non-neoplastic branch; it does not mean that reactive cellular changes are morphologically normal. These urinary labels are not a complete implementation of the Paris System and do not distinguish HGUC from LGUC.

The urinary Diagnosis evaluation compared Reactive_change with Urothelial_carcinoma. Negative was included in source-cohort training but was not a candidate Diagnosis label in that test analysis.

⚙️ Training

LoRA targeted the query, key, value, and output projections of both encoders, with rank 8, alpha 16, dropout 0.10, and no bias training. Training used standard CLIP contrastive loss, AdamW, a learning rate of 1e-5, a batch size of 128, and 100 epochs without a learning-rate scheduler. The checkpoint with the lowest validation loss, at epoch 91, was selected.

Training text prompts used a photo of {label}, with predefined valid single-level and composite hierarchical labels. Training was performed on two NVIDIA RTX A6000 GPUs with random seed 64.

📈 Reported Study Performance

Urinary Cytology

Field-level accuracy on 200 selected fields from 40 independent cases: 20 Reactive change cases and 20 Urothelial carcinoma cases, with five fields per case.

Model	Anomaly	Malignancy	Diagnosis
CLIP ViT-B/16	46.5%	47.0%	44.0%
CLIP ViT-B/16 + LoRA	57.5%	58.0%	60.0%
CYTOLONE-v1.1	53.5%	53.5%	55.0%
CYTOLONE-v2	64.0%	64.0%	72.5%

At the urinary Diagnosis endpoint, with Urothelial carcinoma as the positive class:

Metric	Estimate	95% confidence interval
Sensitivity	99.0%	96.7–100%
Specificity	46.0%	30.0–62.0%
Balanced accuracy	72.5%	64.4–80.5%
Precision	64.7%	48.1–79.4%
F1 score	78.3%	64.7–88.1%

Confidence intervals were estimated using case-clustered bootstrap resampling (20,000 replicates), retaining each case's five fields together.

The model correctly classified 99 of 100 carcinoma fields, but 54 of 100 reactive fields were classified as carcinoma. High sensitivity was accompanied by frequent overcalling of reactive changes. These results are not case-level diagnostic or screening performance estimates.

Cervical Cytology

Field-level accuracy on the independent cervical test cohort from the previous study:

Model	Anomaly	Malignancy	System/Bethesda	Diagnosis
CLIP ViT-B/16	28.2%	33.7%	16.0%	21.7%
CYTOLONE-v1.1	94.7%	93.5%	68.0%	61.0%
CYTOLONE-v2	94.8%	91.2%	59.7%	56.7%

Urinary adaptation preserved coarse-level Anomaly accuracy, while accuracy decreased at the Malignancy, Bethesda, and Diagnosis levels. CYTOLONE-v2 should not be interpreted as a uniform improvement over v1.1 across cervical classification tasks.

⚠️ Intended Use and Limitations

• Intended for research on individual-field cytology classification and cross-specimen adaptation. Outputs require interpretation by qualified professionals and must not be used alone for clinical decisions.
• The model does not automatically locate, rank, or flag suspicious fields, and the study did not evaluate case-level aggregation or autonomous screening.
• The urinary test cohort excluded Negative cases and did not represent the full spectrum of urinary cytology. Atypical and suspicious Paris System categories were not separate targets; grade-specific performance was not evaluated.
• Diagnosis-aware manual field selection introduces selection bias. Routine diagnoses were not subjected to additional study-specific consensus adjudication or uniform histological confirmation.
• Data came from three institutions, but external validation was not performed. Diagnosis distributions differed by institution, so institution-associated effects cannot be excluded.
• Direct microscope-camera images, prospective workflows, reader performance, workload, and clinical utility remain to be validated. Fine-grained cervical performance decreased after urinary adaptation.

The study images are not publicly distributed because of ethical restrictions. Requests for data access should be directed to the corresponding authors of the study.

💻 Application Compatibility

This checkpoint is intended for CYTOLONE 2.0.0, which uses the same v2 classifier for cervical and urinary specimens with different label sets.

CYTOLONE v2 does not use or download the v1/v1.1 classifiers. Legacy classifier entries in Model Management are deletion-only. To use a v1-series classifier, use a compatible CYTOLONE 1.1.x application.

📄 License

Released under CC BY-NC-SA 4.0. Non-commercial reuse and adaptation require attribution and compliance with the license's share-alike terms.

📂 Resources and Citation

• CYTOLONE-v2 study: Kurita Y, et al. A WSI-Independent Inference Framework for Cervical and Urinary Cytology Using Hierarchical Labels. Cytopathology. 2026. Published online October 20, 2026 (Early View). DOI: 10.1111/cyt.70140.
• Original cervical cytology study: Kurita Y, et al. Whole Slide Imaging-Free Supporting Tool for Cytotechnologists in Cervical Cytology. Modern Pathology. 2025;38(10):100817. DOI: 10.1016/j.modpat.2025.100817.
• Application repository: kuri54/CYTOLONE.
• Previous models: CYTOLONE-v1 · CYTOLONE-v1.1.