Currently we only use a 3x1x1 conv adapter in the 3d sam wrapper. This leads to quite slow growth of the receptive field in z, whereas the receptive field in xy is global due to the transformer blocks. This can probably be improved by more convs or a larger depth kernel. Alternative would be the SD Trans approach of SAM-Adapter, which provides a global receptive field across the depth as well.
Currently we only use a 3x1x1 conv adapter in the 3d sam wrapper. This leads to quite slow growth of the receptive field in z, whereas the receptive field in xy is global due to the transformer blocks. This can probably be improved by more convs or a larger depth kernel. Alternative would be the SD Trans approach of SAM-Adapter, which provides a global receptive field across the depth as well.