Fig. 1
Illustration of the Seg2RefineNet framework. (Top) Architecture of the spatio-frequency attention-based network (SFANet) for 2D coronary artery segmentation. The input shows the CCTA slices \(X'_{i}\), obtained from the enhanced 3D CCTA volume \(X'\) and the output shows the initial 2D segmentation masks \(S_{i}\), which are concatenated together to produce the initial 3D segmentation volume S. (Center) 3D refinement pipeline using Attention-GAN. From left to right: The inputs to the generator are the enhanced CCTA volume \(X'\) and the initial segmentation volume S, while the discriminator is trained using \(X'\), S, Y and Ground Truth annotations T. After GAN processing, the refined 3D segmentation volume Y is produced. (Bottom) Detailed generator (G) and discriminator (D) architectures of the 3D Attention-GAN showing downsampling/upsampling blocks, residual connections, and attention mechanisms.
