Table 4.

Model efficiency comparison.

Model	Parameters, million	Inference time (ms/image)	Peak GPU memory (MB)	Training time (min)
Residual U-Net	1.95	2.595	396.6	21.5
Combined augmentation	1.95	2.579	397.4	20.4
Blur augmentation	1.95	2.596	397.3	19.7
Photometric augmentation	1.95	2.605	397.6	19.9
Noise augmentation	1.95	2.612	397.6	19.7
Attention U-Net	1.98	3.020	455.2	21.4
U-Net	1.95	2.620	397.3	20.8

All augmentation variants use the Residual U-Net backbone; augmentation changes training only.