Table 3.

The effect of classifier backbone architecture choices on cell-level performances.

Backbone	#Params	Training time	Brightfield		Brightfield + Hoechst		Brightfield + Fluorescence
			F1	AUROC	F1	AUROC	F1	AUROC
Swin-B38	86.7 M	2.80 h	59.9 ± 0.1	75.8 ± 0.1	66.2 ± 0.1	83.0 ± 0.1	94.6 ± 0.1	99.3 ± 0.1
Swin-S38	48.8 M	2.18 h	59.8 ± 0.3	77.1 ± 0.1	66.2 ± 0.1	83.0 ± 0.1	94.6 ± 0.1	99.3 ± 0.1
ConvNext-L33	196.2 M	1.30 h	61.0 ± 0.1	77.9 ± 0.2	66.4 ± 0.4	83.2 ± 0.2	94.5 ± 0.1	99.3 ± 0.1
ConvNext-B33	87.6 M	0.87 h	60.5 ± 0.4	77.5 ± 0.2	66.0 ± 0.2	83.0 ± 0.2	94.5 ± 0.1	99.3 ± 0.1
ConvNext-S33	49.5 M	0.75 h	60.2 ± 0.1	77.2 ± 0.1	65.8 ± 0.3	82.7 ± 0.2	94.6 ± 0.1	99.3 ± 0.1
EfficientNet-B736	63.8 M	1.30 h	60.3 ± 0.4	77.1 ± 0.2	65.1 ± 0.2	81.9 ± 0.2	94.5 ± 0.2	99.1 ± 0.2
EfficientNet-B436	17.6 M	0.81 h	60.0 ± 0.1	77.3 ± 0.2	65.2 ± 0.3	82.2 ± 0.1	94.4 ± 0.2	99.2 ± 0.1
EfficientNet-B136	6.5 M	0.71 h	59.1 ± 0.2	76.2 ± 0.2	63.4 ± 0.3	80.6 ± 0.1	94.4 ± 0.2	99.2 ± 0.1
DenseNet-20137	18.1 M	1.23 h	60.0 ± 0.4	76.6 ± 0.2	64.6 ± 0.3	81.7 ± 0.3	94.5 ± 0.2	99.3 ± 0.1
DenseNet-16937	12.5 M	0.94 h	59.5 ± 0.2	76.5 ± 0.2	64.8 ± 0.2	81.8 ± 0.1	94.2 ± 0.1	99.2 ± 0.1
DenseNet-12137	7.0 M	0.75 h	59.3 ± 0.4	76.0 ± 0.3	64.2 ± 0.3	81.1 ± 0.2	94.4 ± 0.4	99.1 ± 0.1
ResNet-15230	58.2 M	1.03 h	58.8 ± 0.2	75.7 ± 0.3	63.9 ± 0.3	81.1 ± 0.2	94.5 ± 0.1	99.2 ± 0.1
ResNet-10130	42.5 M	0.78 h	59.1 ± 0.4	75.7 ± 0.1	63.6 ± 0.3	80.8 ± 0.1	94.4 ± 0.2	99.2 ± 0.1
ResNet-5030	23.5 M	0.55 h	59.1 ± 0.2	75.8 ± 0.1	63.8 ± 0.2	80.9 ± 0.1	94.6 ± 0.1	99.3 ± 0.1

Every experiment was conducted using NVIDIA RTX 3090 and Intel(R) Core(TM) i9-9900K CPU @ 3.60 GHz.