表 2. Artificial intelligence models for sperm screening.

用于精子筛选的人工智能模型

Source	Time	Model	Target	Sperm observation method	Performance
CNN: convolutional neural network; SVM: support vector machine; HE: hematoxylin-eosin staining; MAE: mean absolute error.
ILHAN H O, et al.[19]	2020	CNN	Morphology	HE staining	The highest accuracy rate reached 87%.
MIRSKY S K, et al.[23]	2017	SVM	Morphology	Interference microscope	The precision exceeded 90%.
RIORDON J, et al.[24]	2019	CNN	Morphology	HuSHeM, SCIAN dataset	The accuracy rate reached 94.1% (HuSHeM) 　and 62.0% (SCIAN).
JAVADI S, et al.[18]	2019	CNN	Morphology	MHSMA dataset	The accuracy rate reached 84.74% (acrosome), 　83.86% (head), and 94.65% (vacuole).
NOY L, et al.[25]	2023	CNN	DNA	Acridine orange staining	The MAE was 0.05.
McCALLUM C, et al.[21]	2019	CNN	DNA	Acridine orange staining	The accuracy rate for prediction in 10 ms reached 86%.
WANG Y, et al.[20]	2019	Regression model	DNA	Acridine orange staining	The accuracy rate reached 82.7%.
VALIUŠKAITĖ V, et al.[26]	2020	R-CNN	Movement	VISEM database	The accuracy rate reached 91.77% and the MAE was 2.92.
HICKS S A, et al.[27]	2019	CNN	Movement	VISEM database	The lowest MAE was 8.786.
GOODSON S G, et al.[28]	2017	SVM	Movement	CASA analysis	The accuracy rate was 89.9%.
LEE R, et al.[29]	2022	U-Net	Testicular sperm 　sampling	LIVE/DEAD fixable aqua 　dead cell staining	The sensitivity was 86.1% and the F1 score was 85.2%.