基于多尺度特征融合与注意力的肝脏分割方法
Liver segmentation method based on multi-scale feature fusion and attention
冉梅子 1胡小军 2姜晓燕 1范应方 2王航 1王海玲 1高永彬1
作者信息
- 1. 上海工程技术大学电子电气工程学院,上海 201620
- 2. 南方医科大学第五附属医院肝胆外科,广东广州 510000
- 折叠
摘要
由于CT影像对比度低、肝脏形状不规则、相邻器官边界模糊,目前基于卷积神经网络的方法在肝脏分割任务上的表现不佳,尤其是在边界识别和小目标检测方面.基于此,提出一种基于多尺度特征融合与注意力的肝脏分割方法(MFFA UNet).首先,利用多尺度特征融合获取丰富的分割信息,同时使用空间和通道注意力机制捕获全局空间和通道间的关系.其次,通过深度监督模块充分利用中间隐藏层的输出,增强网络的学习能力,加快网络收敛速度.此外,采用一种混合损失函数,以解决类别不平衡的问题,进一步提升模型的分割效能.实验结果表明,所提出的MFFA UNet方法在公共数据集LITS上的表现超越当前主流分割网络,分割结果更接近真实值.
Abstract
Due to the low contrast of CT images,irregular shape of the liver,and blurred boundaries with adjacent organs,the existing methods based on convolutional neural network underperform in liver segmentation tasks,especially for boundary recognition and small object detection.A novel liver segmentation method is proposed based on multi-scale feature fusion and attention,namely MFFA UNet.Multi-scale feature fusion is firstly employed to acquire abundant segmentation details,while spatial and channel attention mechanisms are utilized to capture global spatial and inter-channel relationships.Additionally,a deep supervision module fully leverages the output of intermediate hidden layers,enhancing the learning capability of the network,which in turn accelerates the network's convergence speed.Moreover,a hybrid loss function is adopted to address the issue of class imbalance,further boosting the model's segmentation efficacy.Experimental results demonstrate that the proposed MFFA UNet outperforms the prevailing segmentation networks on the public LITS dataset,producing results that are closer to the ground truth.
关键词
肝脏分割/注意力机制/多尺度特征融合/深度监督/MFFA/UNetKey words
liver segmentation/attention mechanism/multi-scale feature fusion/deep supervision/MFFA UNet引用本文复制引用
基金项目
广州市科技计划(202206010093)
上海市科委社会发展项目科技创新行动计划(21DZ1204900)
出版年
2024
NETL
NSTL
万方数据