Position Encoding UNet architecture. The model consists of a UNet architecture with the patch position as an additional input. Three alternatives of PE-UNet have been explored, Input (in orange), where we input the position as two additional channels before the first layer; Latent (in green), where we input the position as two additional features after the bottleneck; Decision (in red), where we input the position as two additional features before the last convolution. In our experiment, the PE-UNets are composed of five UNet blocks reducing the dimension to 2 × 2 × 1024 at the bottleneck level given a patch of size 32 × 32 × 1.