EdgeAware-CBAM-UNet: A Boundary-Aware U-Net with CBAM and Laplacian Edge-Weighted Loss

EdgeAware-CBAM-UNet is proposed in this work—a medical image segmentation network that combines attention mechanisms with edge-aware loss. The network uses ResNet-18 as its encoder backbone to efficiently extract multi-scale features; integrates CBAM modules into the skip connections to adaptively recalibrate both channel and spatial features; and introduces a Laplacian edge-weighted loss in the objective function to enhance sensitivity to fine vascular branches and tumor margins. We conduct experiments on the CHASE-DB1 retinal vessel dataset and the BraTS2018 brain tumor dataset. The results show that EdgeAware-CBAM-UNet achieves a 2–3% improvement in Dice score over U-Net and U-Net++, while significantly sharpening and preserving boundary contours. Our findings indicate that combining ResNet-18 with an edge-aware loss leads to more accurate and effective segmentation results.