FBB-HFNet: Foreground-background-boundary aware hierarchical fusion network for polyp segmentation

Accurate polyp segmentation is a critical step in the early diagnosis of colorectal cancer. Existing methods commonly suffer from two limitations: insufficient exploitation of intra-layer channel subspaces, and the inability to explicitly handle highly ambiguous boundary pixels under reverse-attention paradigms. To address these issues, we propose FBB-HFNet, a hierarchical feature fusion network that integrates a grouped channel attention enhancement block (GCAEB) and a foreground-background-boundary attention block (FBBAB). The GCAEB partitions encoder features into parallel subspaces and applies independent channel attention to capture complementary channel-subspace cues. The FBBAB explicitly decomposes the predicted probability map into three confidence maps, corresponding to foreground, background, and boundary confusion, from which a deterministic estimator constructs a prototype for each region, and then adaptively aggregates context through a learnable regional prototype attention, effectively alleviating foreground-background confusion near polyp contours. Extensive experiments on five public benchmarks demonstrate that the proposed method achieves competitive or superior performance against current state-of-the-art methods.