New U-Net Variant Combines Hybrid Attention and Dilated Convolution for Brain Tumor Segmentation

Researchers in China have designed a new deep learning model for automatically segmenting brain tumors on magnetic resonance imaging scans, and they report that it achieved Dice coefficients of 87.39 percent on the public BRISC 2025 dataset and 82.84 percent on the Figshare dataset. The work, led by Xiaoshuai Cao of the First Affiliated Hospital of Henan University of Science and Technology, with colleagues at Henan University of Science and Technology in Luoyang, was accepted by Frontiers in Neuroscience on 15 October 2026 after submission on 15 September 2026, and the final formatted version is described as forthcoming.

Automatic segmentation of brain tumors from MRI is regarded as essential for clinical diagnosis and surgical planning, but it remains a difficult task. Tumor margins are often ambiguous, and tumor dimensions can vary markedly, which makes it hard for a single network to capture both local fine details and global contextual cues at the same time. Conventional approaches, the authors note, are hindered by exactly this tension between fine-grained and wide-field information.

To address the problem, the team built a model they call HCSA-RMSDC U-Net, an asymmetric variant of the widely used U-Net architecture. The asymmetry lies in pairing a computationally heavy encoder with a lightweight decoder, concentrating processing capacity on the feature-extraction side of the network while keeping the decoding side lean.

Within the encoder, the researchers incorporated a hybrid channel–spatial attention module, abbreviated HCSA. A distinctive feature of this module is the use of axial decomposed convolution, which is employed to capture directional spatial information. According to the authors, this enhances spatial feature representation for tumor segmentation, sharpening the model’s ability to delineate tumor boundaries.

At the bottleneck layer of the network, the team deployed a second innovation: a residual multi-scale dilated convolution module, or RMSDC. The authors state that this module captures semantic information across varying scales without increasing the spatial resolution, which in turn improves the characterization of tumor heterogeneity.

The two mechanisms are presented as complementary. The HCSA module strengthens spatial feature representation in the encoder, while the RMSDC module extends multi-scale contextual modeling at the bottleneck. The authors argue that the experimental results corroborate the complementary efficacy of the HCSA and RMSDC mechanisms in enhancing feature representation and multi-scale contextual modeling.

Evaluation was performed on two publicly available datasets, BRISC 2025 and Figshare, with the model obtaining Dice coefficients of 87.39 percent and 82.84 percent respectively.

The authors position the framework as a fully automated approach for brain tumor segmentation, with no manual intervention required in the segmentation pipeline. They frame the work as relevant to medical image analysis, multi-scale feature extraction, and hybrid attention methods, listing keywords that include brain tumor segmentation, hybrid channel-spatial attention, magnetic resonance imaging, and multi-scale feature extraction.

As with any newly reported segmentation model, the reported Dice scores come from the authors’ own evaluations on the two named public datasets, and the article is currently at the stage where the final formatted version has not yet been published. The paper carries the standard disclaimer that all claims expressed in the article are solely those of the authors and do not necessarily represent those of their affiliated organizations or the publisher. Readers interested in the technical details can consult the article, which is published open access under a Creative Commons Attribution license and is available through Frontiers in Neuroscience in the Brain Imaging Methods section, with correspondence directed to Xiaoshuai Cao and Jincan Zhang.

Subject of Research: A deep learning model for automatic brain tumor segmentation from MRI using hybrid attention and multi-scale dilated convolution.

Article Title: Asymmetric encoder-decoder U-net with hybrid attention and residual dilated convolution for brain tumor segmentation

Article References: Cao, X., Song, X., Zhang, J., Chen, W., & Du, G. (2026). Asymmetric encoder-decoder U-net with hybrid attention and residual dilated convolution for brain tumor segmentation. Frontiers in Neuroscience, 20, Article 1999080. https://doi.org/10.3389/fnins.2026.1999080

Image Credits: AI Generated

DOI: 10.3389/fnins.2026.1999080

Keywords: brain tumor segmentation, magnetic resonance imaging, U-Net, hybrid channel-spatial attention, dilated convolution, deep learning, medical image analysis, multi-scale feature extraction, Dice coefficient, BRISC 2025, Figshare dataset, encoder-decoder architecture