Table 1 Performance of each U-Net model within the cascading U-Net method. For offline phase, we use a GPU in JURECA High-Performance Computing (HPC) infrastructure at Forschungszentrum Jülich, utilizing a GPU-equipped node with 2 \(\times\) AMD EPYC 7742 processors (2 \(\times\) 64 CPU cores at 2.25 GHz), 512 GB of RAM, and 4 \(\times\) NVIDIA A100 GPUs per node. For online phase, we employ a CPU in a Dell Latitude 7440 Laptop equipped with Gen 13th Intel i7-1365U containing 12 CPU cores at 1.8 GHz and 32 GB RAM.
Model | Number of encoders | Number of filters in the first encoder | Learning rate | Batch size | Number of epochs | Offline time (minutes) | Offline resources | Binary cross-entropy error | Online time per image (seconds) |
|---|---|---|---|---|---|---|---|---|---|
U-Net 1 | 5 | 32 | \(4.4 \times 10^{-4}\) | 2 | 9,269 | 15 | 1 GPU | \(2.7 \times 10^{4}\) | 0.46 |
U-Net 2 | 6 | 64 | \(4.9 \times 10^{-4}\) | 16 | 3,853 | 17 | 1 GPU | \(7.7 \times 10^{2}\) | 1.60 |
U-Net 3 | 5 | 64 | \(6.9 \times 10^{-4}\) | 13 | 2,997 | 16 | 1 GPU | \(2.1 \times 10^{2}\) | 1.22 |
