CG-Net改进的结直肠癌病灶分割算法
Improved CG-Net segmentation algorithm for colorectal cancer lesions
李兰兰 1胡益煌 1王大彪 2徐斌 1李娟3
作者信息
- 1. 福州大学物理与信息工程学院福建省媒体信息智能处理与无线传输重点实验室,福建福州 350108
- 2. 福州大学石油化工学院,福建福州 350108
- 3. 中山大学附属第六医院 内镜外科广东省结直肠盆底疾病研究重点实验室,广东广州 510655
- 折叠
摘要
为解决深度学习分割算法在病灶的细节分割上存在漏判且模型参数量较大不利于实际应用的问题,提出一种基于改进的CG-Net的深度轻量化分割神经网络.在编码块加入改进高效金字塔拆分注意力模块和深度可分离卷积,以学习丰富多尺度全局特征;采用残差思想将注意力模块与编码块结合,提出高效金字塔语境引导模块,帮助网络学习全局和局部特征信息.在中山大学附属第六医院提供的腹部MRI图像数据库的结直肠肿瘤病灶分割实验中,验证了改进模型算法在分割精度和模型轻量化方面的有效性.
Abstract
To solve the problem that the depth learning segmentation algorithm misses judging the detail segmentation of lesions and the large amount of model parameters is not conducive to practical application,a deep lightweight segmentation neural net-work based on improved CG-Net was proposed.The improved efficient pyramid split attention and deep separable convolution were added to the coding block to learn the enrich multi-scale global features.An efficient pyramid context guide module was proposed by combining the attention module and the coding block with the residual idea to help the network learn global and local feature information.In the experiment of colorectal tumor segmentation in the abdominal MRI image database provided by the Sixth Affiliated Hospital of Sun Yat-sen University,the effectiveness of the improved model algorithm in segmentation accuracy and model lightweight is verified.
关键词
深度学习/编码解码网络/轻量级/深度可分离卷积/医学图像分割/注意力机制/结直肠癌Key words
deep learning/encoder-decoder network/lightweight/deep separable convolution/medical image segmentation/at-tention mechanism/colorectal cancer引用本文复制引用
基金项目
福建省自然科学基金项目(2020J01453)
出版年
2024
NETL
NSTL
万方数据