Accurate delineation of grape berry boundaries is essential for phenotypic measurement and growth assessment. This study proposes a multi-source feature fusion network (MFFNet) for instance segmentation of grape berries in dense clusters with frequent overlaps and blurred edges. MFFNet employs two parallel branches for feature extraction: a Swin Transformer backbone to capture hierarchical semantic features and an edge-detection branch that predicts an edge probability map to provide boundary cues. To address the substantial scale variation within a single image, the multilevel semantic features were enhanced using Adaptive Spatial Feature Fusion (ASFF). The edge probability map was introduced twice into the ASFF-enhanced multi-scale features. First, edge cues were injected into the highest-resolution fused feature map to strengthen global boundary awareness across the cluster. Second, during mask generation, edge cues were reintroduced within each candidate instance region to refine local contours and improve the separation in the adhered areas. Experiments on a custom dataset collected in Yinchuan, Ningxia, showed that MFFNet achieved an ${\text{mAP}}_{\text{50}}^{\text{box}}$ of 93.4% and ${\text{mAP}}_{\text{50}}^{\text{mask}}$ of 93.4%, outperforming representative baselines, including Mask2Former and HTC. The proposed model remained stable on images with severe berry overlap and indistinct edges, supporting practical grape growth monitoring.