TECHNOLOGY & EVIDENCE

Make uncertainty visible. Keep expertise in the loop.

YMCA supports abnormal blood cell morphology review through cell-level image analysis, quality checks, human correction and traceable review events.

ENGINEERING PRINCIPLES

An inspectable analysis pipeline.

01

A consistent starting point

Using fixed input images, preprocessing settings and model versions is intended to provide a consistent starting point for professional review.

02

Quality and uncertainty signals

Class probabilities and QC flags bring uncertain, rare or potentially high-risk candidates to the reviewer’s attention. Five uncertainty-related flags block batch acceptance: missing classifier output, low detection confidence, low classifier probability, a small gap between the top two class probabilities and high predictive uncertainty (entropy). Segmentation, overlap, possible multi-cell box and rare-class flags remain visible as advisory signals. These engineering thresholds require clinical calibration.

03

Manual correction and re-analysis

Professionals can correct predicted labels. Original-image annotation adds missed candidates for a subsequent analysis rerun.

04

Planned learning from review

Expert-confirmed corrections can inform future datasets. The walkthrough shows morphology explanations and reference-cell comparisons; clinical calibration of these aids and expanded model governance remain development priorities.

SYSTEM COMPONENTS

How the system supports review.

Each stage produces an output that is stored with the case, so the reviewer and the next stage can inspect it.

01 / DETECTION

Candidate detection

A detection model proposes candidate cells and stores each box in original-image coordinates, so every result traces back to its source location.

02 / CELL ANALYSIS

Cell-level image analysis

Each candidate is cropped and segmented into a focused cell image. The segmentation output is kept with the record and flagged when it may be unreliable.

03 / CLASSIFICATION

16-category morphology pre-classification

The classifier stores a probability for every one of the 16 categories, not only the top label, so uncertainty can be checked later. The categories include mature and immature cell types, labels suggestive of abnormal morphology, and artifacts.

04 / ASSISTANCE

Review assistant

The assistant reads the same stored cell records and comparisons as the interface. A review change it proposes goes through the same confirmation step and review history as a manual decision.

The prototype supports image-based research review. Clinical integration and validation remain development priorities.

CURRENT MODEL EVALUATION

Initial evidence, clearly scoped.

16-category classifier · Fixed image-level held-out test

3,212Held-out test images
16Morphology categories
0.8164Macro-F1 score
96.23%Overall classification accuracy

What these results mean

These results cover 3,212 held-out cell images from the team’s prepared research dataset, using a 16-category morphology classifier. The split is image-level; patient or slide overlap between training and testing has not been ruled out. Patient-level, slide-level, external and end-to-end clinical performance remain to be validated.

The 16 categories support abnormal morphology review; they are not 16 exclusively abnormal cell types. The label set also includes neutrophils, monocytes, mature lymphocytes, artifacts and erythroid cells. Some rare-class test subsets contain only 1–6 images. Their estimates are unstable, and high recall in these subsets must be interpreted alongside false positives and sample size.

NEXT VALIDATION PRIORITIES

Measure the workflow, as well as the model.

Our planned studies compare traditional review with AI-assisted review, with professional oversight throughout.

Clinical usability

Review time, workload and user feedback. A ≥30% reduction in median review time per slide is a study target, not a demonstrated result.

Agreement and safety

Inter-reader agreement, class-specific recall, correction patterns and repeat review.

External generalization

Patient- and slide-level testing across hospitals, staining conditions and imaging devices.

DATA & MODEL GOVERNANCE

Expert feedback. Controlled updates.

The planned data cycle keeps clinical source data within hospital governance. Only authorized, de-identified data enter controlled training environments.

  1. Professional review
  2. Expert confirmation & data approval
  3. Offline training
  4. Independent validation
  5. Versioned deployment

Production models are not intended to learn autonomously during clinical use. YMCA has not obtained medical-device clearance; clinical deployment and regulatory pathways remain under development.