| 方文胜,杨庆苒,徐争元.一种轻量化多尺度注意力引导的肝脏及肿瘤CT分割网络[J].井冈山大学自然版,2025,46(6):63-70 |
| 一种轻量化多尺度注意力引导的肝脏及肿瘤CT分割网络 |
| A lightweight multi-scale attention-guided network for liver and tumor segmentation in CT images |
| 投稿时间：2025-06-07 修订日期：2025-07-12 |
| DOI：10.3969/j.issn.1674-8085.2025.06.007 |
| 中文关键词: 肝脏CT分割 肝肿瘤分割 多尺度空洞卷积 多尺度自适应层级注意力 |
| 英文关键词: liver CT segmentation liver tumor segmentation multi-scale dilated convolution multi-scale adaptive hierarchical attention |
| 基金项目:安徽省高等学校科学研究项目(自然科学类)重点项目(2024AH051916); 皖南医学院校中青年科研基金项目(WK202203) |
|
| 摘要点击次数: 66 |
| 全文下载次数: 53 |
| 中文摘要: |
| 针对肝脏及其肿瘤CT图像分割中存在的边界模糊、多尺度目标检测及低对比度等问题，本研究提出一种轻量化多尺度注意力引导的U-Net网络（MSAU-Net）。所提模型引入深度可分离卷积（DWConv），在降低模型参数量与计算复杂度，同时保持特征表达能力；提出多尺度空洞卷积模块（MSDConv）并行提取不同尺度特征以增强小目标检测；在瓶颈阶段结合多尺度自适应层级注意力机制（MSA-HAM），实现跨空间、深度与尺度维度的特征加权；并采用注意力引导的跳跃连接（AG）提升目标与背景分割对比度。实验结果表明，在SLIVER07与Li TS17数据集上，MSAU-Net肝脏和肿瘤分割任务中Dice系数分别达到0.977 2、0.965 7和0.876 2，性能优于当前主流方法。该模型同时具备高精度与低复杂度，适用于资源受限环境下的智能医学影像分割任务。 |
| 英文摘要: |
| To address the challenges of boundary ambiguity, multi-scale target variation, and low contrast in liver and tumor CT image segmentation, this paper presents a lightweight multi-scale attention-guided network termed MSAU-Net. Specifically, depthwise separable convolution（DWConv） is utilized to reduce the model parameters and computational complexity while preserving feature representation capability. A multi-scale dilated convolution module（MSDConv） is designed to extract the features across multiple receptive fields in parallel,thus improving the detection of small targets. Additionally, a multi-scale adaptive hierarchical attention module（MSA-HAM） is integrated into the bottleneck to reweight the features across spatial and channel dimensions.Attention-guided skip connections（AG） are further employed to enhance the contrast between the foreground and background regions. The experimental results on the SLIVER07 and LiTS17 datasets demonstrate that MSAU-Net achieves Dice coefficients of 0.977 2 and 0.965 7 for liver segmentation and 0.876 2 for tumor segmentation, surpassing the existing mainstream methods. The proposed model offers a favorable balance between the segmentation accuracy and the computational efficiency, which is suitable for deployment in resource-constrained environments for the intelligent medical image segmentation. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
