Table 7.

Results of the ablation study with the combined segmentation on the test data set: For this evaluation, a reduced image resolution of 320 × 256 pixels was used. The results highlight the marginal impact of the different architecture choices (U-Net vs. Feature Pyramid Network (FPN)) as well as the different backbones.

Combined U-Net	PQ	F1	Precision	Recall	Cat. Acc
ResNet34	0.7863	0.9457	0.9559	0.9358	0.9694
Incep.-RN-v2	0.7685	0.9326	0.9501	0.9157	0.9674
EfficientNet-B5	0.7768	0.9404	0.9471	0.9337	0.9692
Combined FPN	PQ	F1	Precision	Recall	Cat. Acc
ResNet34	0.7824	0.9442	0.9556	0.9332	0.9709
Incep.-RN-v2	0.7784	0.9414	0.9511	0.9319	0.9700
EfficientNet-B5	0.7861	0.9451	0.9556	0.9347	0.9709