specifically, the RD3B contained a few 3D deformable convolution layers plus a shortcut. The multiscale spatial branch contained a multiscale spatial block and a concatenate operation, which could learn multiscale https://www.sayspatial.com/