Table 3.

Cell classification performances of hematologists and the deep learning model in the development cohort.

Cell class	Precision/recall/APa (%)
	Hematologist-1 (V10)b	Hematologist-2 (V8)b	Hematologist-3 (V6)b	Artificial intelligence modelc
Cellsd (n=17,319)	N/Ae	N/Ae	N/Ae	55.8/85.6/67.4
Erythroid (n=2967)	87.6/92.3/N/Ae	88.0/94.1/N/A	89.1/91.4/N/A	85.0/84.5/49.1
Blasts (n=4063)	91.0/85.2/N/A	79.1/88.2/N/A	87.5/88.5/N/A	86.5/80.7/50.2
Myeloid (n=2506)	79.1/94.2/N/A	92.0/93.5/N/A	93.8/80.0/N/A	94.0/76.4/49.5
Lymphoid (n=1619)	59.0/78.4/N/A	67.1/79.7/N/A	61.2/71.9/N/A	74.0/58.9/21.9
Plasma cells (n=600)	84.0/92.6/N/A	82.3/96.7/N/A	84.9/81.4/N/A	53.4/74.1/30.0
Monocyte (n=192)	25.9/88.7/N/A	65.7/37.5/N/A	40.2/64.5/NA	57.4/30.0/6.1
Megakaryocyte (n=42)	84.1/97.0/N/A	52.9/61.5/N/A	96.8/100.0/N/A	71.0/56.4/19.0
Unable to identify (n=5330)	86.5/78.5/N/A	82.3/77.5/N/A	83.9/93.5/N/A	60.9/86.1/25.1

^aAP: average precision based on the area under the precision-recall curve.

^bThe abbreviation V(X) denotes a visiting staff member with (X) years of practice experience.

^cAll results were based on 6-fold cross-validation.

^dBounding box prediction performance regardless of the classifications (only for the deep learning model).

^eNot available.