Classification
→
One label
“abnormal”
“abnormal”
Summarizes an entire image into one category, but does not show where the relevant cells are.
AI-ASSISTED CERVICAL CYTOLOGY
Turn every pixel into a reviewable cell boundary.
An interactive presentation of a U-Net based workflow that identifies cell regions in Pap-smear microscopy images, overlays predicted masks, and converts segmentation into measurable morphology.
256 × 256Model input
1-class maskPixel-level output
U-NetEncoder–decoder
Powered with ❤️ by Solicitous ®
X 184.2 µm
Y 093.8 µm
01 / 18
THE IMPORTANT DISTINCTION
This is not merely “cancer / no cancer.”
The model produces a spatial map. Every output pixel answers a more useful question: does this location belong to the target cell region?
Semantic segmentation
→
Produces a binary mask, enabling overlays, area measurements, morphology, quality checks, and human review.
PIXEL-TO-INSIGHT PIPELINE
See the image transform, one decision layer at a time.
STAGE 01 · ACQUISITION
Capture the complete field of view
A microscope-mounted camera supplies a color Pap-smear image. The full frame remains available for review while the analysis pathway creates a standardized model input.
- Preserve original image
- Assign case and slide identifiers
- Record magnification and capture source
SMART SECTION · MORPHOLOGY LAB
Segmentation becomes useful when it becomes measurable.
This interactive panel demonstrates how a predicted mask can drive practical cell-level measurements. Values are presentation simulations derived from the selected sample, not clinical results.
28.4 µm
17.9 µm
SEGMENTATION PARAMETERSLIVE
DATASET ATLAS
Broad cytology context, focused segmentation experiment.
The source microscopy database contains five cervical-cell categories. This experiment prepares image–mask pairs and focuses its demonstrated segmentation training on the Metaplastic category.
DyskeratoticAbnormal squamous morphology
Context classKoilocytoticCharacteristic perinuclear change
Context classMetaplasticTarget class in this segmentation demo
Model focusParabasalSmall immature squamous cells
Context classSuperficial–IntermediateMature epithelial cell patterns
Context classTRAINING IMAGE ATLASAugmented samples
PIXEL LABEL ATLASBinary masks
MODEL ANATOMY
A seven-level encoder–decoder that compresses context, then restores detail.
ENCODERWhat is present?
AI
Latent representationDECODERWhere is it?
SMART SECTION · HUMAN-IN-THE-LOOP QA
Compare prediction and reference without hiding either one.
Move the divider to inspect the full image. No scientific media is cropped.
↔
ORIGINAL MICROSCOPY
EXPERT MASKS
Inspect missed regions, over-segmentation, merged objects and weak edges.
The model assists review; it should not remove expert oversight.
Preserve a traceable record of the input, output and operator response.
TRAINING EVIDENCE
Show the curves, the configuration and the limits.
The presentation avoids inventing clinical performance. It displays the project’s recorded training artifacts and explains what they do—and do not—establish.
LOSS CURVEEarly stopping near epoch 35
ACCURACY CURVETrain vs validation
METRIC TRACEBinary accuracy history
Research evidence ≠ clinical validation
Training curves indicate optimization behavior on the prepared dataset. They do not establish population-level diagnostic accuracy, device performance, or regulatory readiness.
- Single demonstrated target category
- Dataset-specific preprocessing
- External validation still required
- Expert review remains essential
RESULTS THEATRE
Original field, predicted mask and merged inspection view.
Original microscopy test set
FROM MODEL TO INSTRUMENT WORKFLOW
A practical presentation layer for microscope and laboratory ecosystems.
01Digital microscopeCamera-captured slide field
IMAGE
02CytoVision AISegmentation + morphology
MASK
03Expert reviewAccept, refine, annotate
DATA
04Structured reportImages, masks and metrics
Pair microscope hardware with a visual AI review experience rather than presenting inference as a hidden black box.
Use clear overlays, QC prompts and measurement panels that remain understandable for mixed-skill laboratory teams.
The presentation concept can grow toward multi-cell detection, annotation, batch review, reports and local integrations.
CYTOVISION AIMake the segmentation visible.
Return to workstation ↑
Make the segmentation visible.
Make the result reviewable.
Research presentation for AI-assisted cervical cytology workflows.