Table 1.

A performance comparison between the U-Net and Detectron2 on the test data set, using different combinations of backbone, batch size, optimizer and loss functions (bold represent the best performance).

Model Name	Backbone	Batch Size	Optimizer	Loss Function	Metrics
					Accuracy	Sensitivity	Specificity	Dice	AJI
U-Net	ResNet50	2	SGD	BCE	0.97135	0.81653	0.97349	0.51159	0.35148
		4			0.97729	0.66402	0.98229	0.5015	0.34194
		8			0.98119	0.19418	0.99565	0.25376	0.14849
		2	Adam	BCE	0.98922	0.65317	0.99571	0.67643	0.5301
		4			0.98933	0.65703	0.99585	0.67773	0.53082
		8			0.98904	0.65452	0.99551	0.66949	0.52215
	ResNet101	2	SGD	BCE	0.97202	0.80081	0.97434	0.51812	0.35626
		4			0.97635	0.71966	0.98038	0.52575	0.36193
		8			0.98066	0.20058	0.99493	0.25703	0.15036
		2	Adam	BCE	0.98902	0.65415	0.99538	0.67106	0.52394
		4			0.98903	0.66182	0.99542	0.67364	0.52681
		8			0.98939	0.67254	0.99584	0.68844	0.53922
Detectron2	ResNet50	2	SGD	BCE+L1	0.98795	0.63321	0.99509	0.6571	0.50428
		4			0.98823	0.58092	0.99632	0.63354	0.48617
		8			0.98816	0.57355	0.99629	0.62887	0.48037
		2	Adam	BCE+L1	0.98811	0.63597	0.99514	0.65672	0.50619
		4			0.98792	0.57015	0.99616	0.61928	0.47275
		8			0.98823	0.58092	0.99632	0.63354	0.48617
	ResNet101	2	SGD	BCE+L1	0.9881	0.62078	0.99563	0.65472	0.50358
		4			0.98778	0.5791	0.99607	0.62088	0.47237
		8			0.98788	0.58846	0.99609	0.63493	0.48353
		2	Adam	BCE+L1	0.98828	0.62597	0.99563	0.65985	0.50696
		4			0.98817	0.59231	0.9963	0.63697	0.48881
		8			0.98815	0.59255	0.99622	0.63644	0.48773