Table 1.

Comparison table of studies associated with the cell counting procedure

Paper	AI Architecture	Input data type	Training data size	CNN function	Annotation limit	Overall accuracy
Khan et al. (2016) [33]	Custom CNN architecure [39]	Dark-field time-lapse image sequence	265 sequence/ 148.993 frames	Automatically count number of cells	Up to 5 cells stage	± 92.18%
Ng and McAuley et al. (2018) [35]	Modified ResNet model [40]	EmbryoScope time lapse videos	1309 videos/ 191.449 frames	Automatically detect embryo developmental stage	Up to 4+ cells stage	± 87%
Malmsten et al. (2019) [36]	InceptionV3 model [41]	Time-lapse microscopy image	47.584 images	Predict/detect cell division	Up to 4-cell stage	± 90.77%
Liu et al. (2019) [42]	Multi-task deep learning with dynamic programming (MTDL-DP)	Time-lapse microscopy video	Extracted 59.500 frames in total	Classify embryo development stages	From initialization (tStart) to 4+ cells (t4+)	± 86%
Leahy et al. (2020) [37]	ResNeXt101 [43]	Embryoscope time-lapse images	73 embryos with 23.850 labels	Multiple function pipeline, one of which is detecting embryo developmental stage	From start of incubation to finish	± 87.9%%
Dirvanauskas et al. (2019) [38]	AlexNet [39] + 2nd classifier	Miri TL-captured images	3000 images from 6 embryo sets + 600 images of embryo fragmentation	Classify the development stage of an embryo	Classify for 1 cell, 2 cells, 4 cells, 8 cells and No Embryo	± 97.5%%
Lau et al. (2019) [44]	RPN + ResNet-50 [40]	EmbryoScope-captured videos	1309 time-lapse videos extracted by frames	Locate cell and classify development stage	Classify for tStart, tPnf, t2, t3, t4, t4+	± 90%
Raudonis et al. (2019) [45]	Comparing VGG [46] and AlexNet [39]	Miri TL-captured images	300 TL sequences for a totl of 114793 frames	Locate cell and classify development stage	Classify for 1 cell, 2 cells, 3 cells, 4 cells, > 4 cells	VGG: ± 93.6% AlexNet: ± 92.7%