Skip to content
Open access

BRF-net: bidirectional spatial–channel fusion and frequency-aware refinement for medical image segmentation

Nov 2026 · The Visual Computer · Vol 42 · 0 citations · 53 references

Abstract

Accurate dense prediction in medical images remains challenging because anatomical targets often exhibit large-scale variation, low-contrast boundaries, heterogeneous textures, and fine structural details that are easily degraded during feature extraction. Although CNNs provide strong local inductive biases and Transformers capture long-range dependencies, existing CNN, Transformer, and hybrid CNN–Transformer segmentation models still struggle to simultaneously achieve scale-adaptive representation, spatial–semantic consistency, and boundary-sensitive detail preservation. To address these challenges, we propose BRF-Net, a boundary-aware hybrid CNN–Transformer framework for medical image segmentation. From a visual computing perspective, BRF-Net is designed as a dense visual representation architecture for segmentation under scale variation, semantic ambiguity, and boundary degradation, with medical images serving as a stringent testbed for weak boundaries, small structures, and heterogeneous appearances. Specifically, an adaptive gated multi-scale (AGMS) block performs pixel-wise receptive-field selection to adapt to variable target morphology; a bidirectional refinement and fusion (BRF) Attention Block enables reciprocal interaction between spatial window attention and channel self-attention; and a patch-wise Fourier feed-forward network (PF-FFN) introduces learnable local spectral filtering to preserve boundary-sensitive high-frequency information. Extensive experiments on eight public benchmarks covering CT, MRI, ultrasound, dermoscopy, endoscopy, and microscopy show that BRF-Net achieves strong and competitive performance across diverse medical image segmentation tasks. It obtains the best average results on the six binary datasets and the Synapse multi-organ dataset, while remaining highly competitive on ACDC. Together with ablation studies, statistical analyses, qualitative comparisons, Grad-CAM visualization, and computational efficiency evaluation, these results indicate a favorable balance between segmentation accuracy, boundary preservation, and model complexity under the present benchmark evaluation. The source code, fixed data splits, preprocessing and configuration files, and training, inference, evaluation, and statistical testing scripts are released on GitHub and Zenodo with DOI: https://doi.org/10.5281/zenodo.19129179.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.