UAC-Net: a unified multi-task network with aggregated contextual attention and cross-layer fusion for brain tumor segmentation and classification
- 1College of Information Engineering, Henan University of Science and Technology
- 2The First Affiliated Hospital, and College of Clinical Medicine of Henan University of Science and Technology
Abstract
Introduction: Accurate segmentation and classification of brain tumors from magnetic resonance imaging (MRI) are critical for clinical diagnosis and treatment planning. However, existing methods typically treat these two tasks independently, failing to exploit their inherent complementarity, and often struggle to capture multi-scale contextual information essential for identifying small or irregular lesions.
Methods: To address these limitations, this paper proposes UAC-Net, a unified multi-task learning framework built upon a U-Net backbone that jointly optimizes tumor segmentation and classification through a shared encoder. Two core modules are introduced: (1) the Aggregated Contextual Attention (ACA) module, which enhances dynamic perception of critical lesion regions by combining local sliding-window attention with global downsampling attention across all encoder levels; and (2) the Adjacent Layer Multi-scale Feature Fusion (ALMFF) module, which enables cross-level feature interaction to enrich multi-scale representations. In the segmentation branch, ALMFF-fused features are progressively upsampled to achieve precise lesion localization; in the classification branch, ACA-enhanced deep features are aggregated via global average pooling to yield robust semantic representations for tumor typing.
Results: Experiments on two publicly available brain tumor MRI datasets demonstrate that UAC-Net achieves competitive and consistent performance compared with several representative methods. Specifically, UAC-Net obtains a classification accuracy of 98.53% and a Dice coefficient of 81.39% on the Figshare dataset, and achieves 99.07% classification accuracy and 86.61% Dice coefficient on the BRISC 2025 dataset.
Discussion: These findings suggest that integrating contextual attention and cross-layer feature fusion within a unified multi-task framework is a promising approach for joint brain tumor segmentation and classification.
Related
- Hybrid feature pyramid attention and capsule network (FpaNet+CapsNet) for enhanced spatial and contextual feature extraction for improved brain tumor segmentation
- TiLCA-Net: nuclei segmentation and classification in pathological images via task-specialized attention and efficient decoding
- COVLIAS 3.5: integration of attention-based segmentation technique with fuzzy dilated convolutional neural networks for improved classification of chest X-ray scans for multiclass pneumonia diagnosis
- Traffic surveillance video-based a unified road accident detection, classification, and vehicle tracking framework using an optimized spatio-temporal attention mechanism
- A collaborative hybrid network with aggregated attention for pulmonary nodule detection
BibTeX
@article{cai2026uacnet,
title = {{UAC-Net: a unified multi-task network with aggregated contextual attention and cross-layer fusion for brain tumor segmentation and classification}},
author = {Chuanqi Cai and Jincan Zhang and Dongyan Zhang and Wenna Chen and Xinghua Tan and Ganqin Du and Hongwei Jiang},
journal = {Frontiers in Neuroscience},
year = {2026},
month = oct,
doi = {10.3389/fnins.2026.1986322},
url = {https://www.frontiersin.org/articles/10.3389/fnins.2026.1986322/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - UAC-Net: a unified multi-task network with aggregated contextual attention and cross-layer fusion for brain tumor segmentation and classification
AU - Chuanqi Cai
AU - Jincan Zhang
AU - Dongyan Zhang
AU - Wenna Chen
AU - Xinghua Tan
AU - Ganqin Du
AU - Hongwei Jiang
PY - 2026
DA - 2026/10/20
JO - Frontiers in Neuroscience
DO - 10.3389/fnins.2026.1986322
UR - https://www.frontiersin.org/articles/10.3389/fnins.2026.1986322/full
AB - Introduction: Accurate segmentation and classification of brain tumors from magnetic resonance imaging (MRI) are critical for clinical diagnosis and treatment planning. However, existing methods typically treat these two tasks independently, failing to exploit their inherent complementarity, and often struggle to capture multi-scale contextual information essential for identifying small or irregular lesions. Methods: To address these limitations, this paper proposes UAC-Net, a unified multi-task learning framework built upon a U-Net backbone that jointly optimizes tumor segmentation and classification through a shared encoder. Two core modules are introduced: (1) the Aggregated Contextual Attention (ACA) module, which enhances dynamic perception of critical lesion regions by combining local sliding-window attention with global downsampling attention across all encoder levels; and (2) the Adjacent Layer Multi-scale Feature Fusion (ALMFF) module, which enables cross-level feature interaction to enrich multi-scale representations. In the segmentation branch, ALMFF-fused features are progressively upsampled to achieve precise lesion localization; in the classification branch, ACA-enhanced deep features are aggregated via global average pooling to yield robust semantic representations for tumor typing. Results: Experiments on two publicly available brain tumor MRI datasets demonstrate that UAC-Net achieves competitive and consistent performance compared with several representative methods. Specifically, UAC-Net obtains a classification accuracy of 98.53% and a Dice coefficient of 81.39% on the Figshare dataset, and achieves 99.07% classification accuracy and 86.61% Dice coefficient on the BRISC 2025 dataset. Discussion: These findings suggest that integrating contextual attention and cross-layer feature fusion within a unified multi-task framework is a promising approach for joint brain tumor segmentation and classification.
SN - 1662-453X
SN - 1662-4548
ER -