Blood cell review
CellReview shows the blood cell images with the most uncertain predictions first. Using saved predictions and known dataset labels, this experiment tests whether that order finds more disagreements between predictions and labels than reviewing the same number of images at random.
Research prototype · No clinical validation
disagreements found
versus on average at random.
Review queue
Highest uncertainty first.
Use ← and → to browse. Progress stays in this browser and never changes labels.
Classification and review results
Held-out test results, evaluated after fixing the method. The data still come from a single source.
Disagreements found by review budget
Disagreements found with the dataset’s original labels.
P2.5–P97.5 describes variation between random selections of these same images. It is not an interval of clinical effectiveness. Review budgets are rounded up.
How to read this comparison
This is a retrospective simulation: PBC labels are consulted after ranking. Your review progress does not change these metrics.
Even a prediction with very concentrated scores can be wrong. This queue does not guarantee that every disagreement will be found.
Results by cell category
Confusion matrix
Rows: reference label. Columns: prediction. The diagonal contains matches.
Method and evaluation
Experiment
In PBC, labels are already known, so disagreements could be filtered directly. Here they are used to evaluate a ranking strategy for new images without labels, when only some can be reviewed. The queue uses model scores; references are consulted afterwards.
- Data. Original peripheral blood images from PBC, covering eight categories. Detected duplicate groups were kept together and conflicting labels excluded.
- Features. An ImageNet-pretrained ResNet18, kept frozen, produces 512 numerical features per image.
- Classifier. A scaler and logistic regression were fitted on training images only.
- Review priority. Normalized entropy of the eight scores, highest first. Each preset budget is compared with 1,000 random selections of the same size.
Frozen ResNet18 features and logistic regression were chosen as a manageable CPU baseline, keeping feature extraction separate from classifier fitting. The aim was to test review prioritisation with a reproducible reference before considering a more complex model.
The classifier was fitted on 10,246 training images. A separate set of 3,414 development validation images was used to assess the method before the final test. This demo shows only the 3,414 held-out test images, evaluated after fixing the method. Development results remain available in the technical note and validation data.
The application displays saved predictions. It runs no inference or training and does not accept new images. Review marks stay in your browser and do not update labels or results.
Limitations
- One source, with specific sample preparation and image capture.
- Patient identifiers are unavailable. Duplicate grouping does not establish patient-level separation.
- This final test is an internal evaluation; it does not establish generalization to another hospital.
- Scores are uncalibrated: uncertainty is not the probability of an error.
- A disagreement with a dataset label is not a confirmed diagnostic error.
No external-hospital, prospective or human reviewer study has been performed. Clinical benefit and time savings have not been measured. References are available in the downloaded data; hiding them on screen does not constitute a blinded study.
Data and author
CellReview is a research project by David Méndez Merino.
Images: Acevedo, Merino, Alférez, Molina, Boldú and Rodellar (2020), PBC · Mendeley Data v1 · DOI 10.17632/snkd93bnjr.1, licensed under CC BY 4.0. Original JPEGs are redistributed unchanged; the interface scales them for viewing. Training preprocessing used resizing, a central crop and ImageNet normalization.
Feature extractor: ResNet18 from ONNX Model Zoo, pretrained on ImageNet. The saved experiment ran on a local CPU.
Evaluation code and reproduction steps: recalculate metrics and review comparisons from the saved predictions, without retraining.
Technical details and files
The browser checks data and original images against the SHA-256 manifest before displaying them. This detects changes relative to that manifest; it does not authenticate the publisher or establish scientific validity.
Progress is stored in this browser for these saved predictions. Review actions are not sent to a server. The application uses no external scripts, fonts or analytics.
Technical note · Data notice · Collection catalogue · Final-test data · Development-validation data · File integrity manifest
Displayed numbers are rounded. The collection files retain the saved values.