Fig. 5.

Architecture of the segmentation head in the 3D EfficientDet network. The $P_3$-$P_7$ output features from the EfficientNetB0 backbone (see Fig. 2) are first convolved to a common channel dimension of 48 with $1\times 1\times 1$ filters (in white), then iteratively added with outputs from lower levels by fast normalized feature fusion (Eq. 6). The upscaling (purple blocks) is done by a nearest-neighbor resize followed by a $3\times 3\times 1$ anti-aliasing convolution. All convolutions are performed depth-wise and are followed by batch norm and a swish activation function, apart from the last block, which is a single convolution with a $1\times 1\times 1$ kernel and a sigmoid activation function