AttentionUNetConfig

Configuration for Attention U-Net.

Extends the standard U-Net architecture (Ronneberger et al., 2015) by adding soft attention gates on skip connections. Each gate computes a spatial attention map from the skip feature and the gating signal from the decoder, suppressing irrelevant activations before concatenation.

Architecture overview: Encoder: depth × (2×Conv3x3-BN-ReLU + MaxPool2x2) Bottleneck: 2×Conv3x3-BN-ReLU Decoder: depth × (Upsample/ConvTranspose + AttentionGate + Cat + 2×Conv3x3-BN-ReLU) Head: Conv1x1 → num_classes

Args: num_classes: Number of output segmentation classes. in_channels: Number of input image channels. base_channels: Feature channels at the first encoder stage. Doubles at each depth level. depth: Number of encoder/decoder stages (excluding bottleneck). bilinear: If True, use bilinear upsampling; otherwise ConvTranspose2d. deep_supervision: Attach a 1x1 classifier to every decoder level, resample each to the input resolution, and fuse the stack with a final 1x1 convolution — the dsv in the reference's unet_CT_multi_att_dsv model. Despite the name this is not a training-only auxiliary loss: the prediction itself comes out of the fusion, so switching it on changes inference too. Defaults off so existing checkpoints keep loading.