DETECT.CLASSIFY.COUNT.

Computer-aided hematology from a single smear. Localize, name, and quantify cells in one field of view.

001 — THE EYE SAVER

Counting 500 cells by hand

is a great way to lose your sanity.

We taught YOLO to do the staring,

so clinicians don't have to.

85.4% mAP@50 · 36 held-out fields · research prototype

Wright–Giemsa · 100x

01 · Detect
Localize every cell in the field.
02 · Classify
RBC, WBC, or platelet at live gates.
03 · Count
Per-class totals for this field of view.

002 — THE SPECIMEN

Place thespecimen.

Drop a microscopic field

JPG, PNG, BMP · Wright–Giemsa fields work best

003 — THE HELD-OUT FIELD

The curve,then the gate.

Detector quality is measured as COCO mAP on 36 held-out fields (471 instances). The lab does not count on that curve. It infers at 0.40, then keeps RBC ≥ 0.60, WBC ≥ 0.40, platelets ≥ 0.40.

mAP@50

85.4%

baseline 81.81%

mAP@50-95

60.07%

baseline 56.47%

Precision

82.46%

baseline 71.74%

+3.6 mAP@50 vs our own baseline after label correction

Per-class mAP@50

WBC

96.9%

baseline 96.9%

RBC

85.4%

baseline 83.4%

Platelets

73.9%

baseline 65.1%

COCO mAP at conf=0.001. Counting in the lab uses live class gates, not this curve. Platelets remain the hard class. Research prototype · not a diagnostic device.