Fig. 4: Cross-scale consensus fusion and pyramid refinement decoder (CSCF-PRD): integrating global-local features and reconstructing dual outputs.
From: GLANCE: continuous global-local exchange with consensus fusion for robust nodule segmentation

Global and local features are merged at each scale via CSCF and progressively upsampled through the PRD. The final dual heads produce both a segmentation map and a detection heatmap, enabling precise and interpretable multi-scale nodule analysis.