| 摘 要: 针对肝脏医学影像分割中存在的边缘模糊、噪声干扰等问题,提出了一种基于双支路特征融合的分割网络。其中,编码器采用并行策略,一路为改进的卷积神经网络(Convolutional Neural Network,CNN)支路,采用密集提取特征的集成模块,提供了丰富的局部信息;另一路为Transformer支路,负责提供空间维度上的全局信息。将两支路相同层级的特征进行交互式融合,提高了特征信息的表达能力。最终,通过解码得到分割预测结果。在肝脏肿瘤分割挑战赛采用的公开数据集(Liver Tumor Segmentation Challenge,LiTS)上的实验表明,该网络在肝脏和肝脏肿瘤的分割上分别达到了95.64%和86.89%的精度,实现了较好的分割效果。 |
| 关键词: 医学图像分割;特征融合;CNN;Transformer;注意力机制 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(61971272) |
|
| Dual-Channel Feature Fusion Network for Liver and Liver Tumor Segmentation |
|
ZHANG Mengying, HE Lifeng, ZHU Fen, SUN Shuang
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, X'i an 710021, China)
zz13060325591@163.com; helifeng@ist.aich-i pu.ac.jp; 1633733405@qq.com; 973615721@qq.com
|
| Abstract: To address the challenges of blurred boundaries and noise interference in liver medical image segmentation, this study proposes a segmentation network based on dua-l channel feature fusion. The encoder adopts a parallel strategy: one channel is an enhanced Convolutional Neural Network (CNN) pathway equipped with a dense feature extraction module to capture rich local information, while the other channel employs a Transformer pathway to extract global spatial context. Interactive fusion of features from corresponding hierarchical levels of both channels enhances feature representation capabilities. The decoded output generates the final segmentation prediction. Experiments on the publicly available dataset used in the Liver Tumor Segmentation Challenge (LiTS) demonstrate that the network achieves segmentation accuracies of 95.64% for liver segmentation and 86.89% for liver tumor segmentation, resulting in improved segmentation performance. |
| Keywords: medical image segmentation; feature fusion; CNN; Transformer; attention mechanism |
