Fig. 4

The architecture of CGS-Net for semantic segmentation using MiT-B1 encoders²⁸, which takes into account both details and context. The two input images with the same dimensions are the level-2 high-magnification (represented by blue), providing a focused view, and the level-3 low-magnification (represented by red), providing a wider context. The encoder for the detail shown in blue and the encoder for the context shown in red. The center of the figure is the cross attention blocks, where the context feature tensors serve as the query input and the detail feature tensors are the keys and values input. The right side of the figure shows the decoder, which extends the standard U-Net decoder by adding two additional double convolutions and transposed convolutions.