欢迎访问《图学学报》 分享到:

图学学报 ›› 2023, Vol. 44 ›› Issue (2): 225-232.DOI: 10.11996/JG.j.2095-302X.2023020225

• 图像处理与计算机视觉 • 上一篇    下一篇

动态平衡多尺度特征融合的结直肠息肉分割

  

  1. 1. 桂林理工大学信息科学与工程学院,广西 桂林 541004;  

    2. 广西嵌入式技术与智能系统重点实验室,广西 桂林 541004

  • 出版日期:2023-04-30 发布日期:2023-05-01
  • 基金资助:
    国家自然科学基金项目(62166012) 

Dynamic balanced multi-scale feature fusion for colorectal polyp segmentation 

  1. 1. School of Information Science and Engineering, Guilin University of Technology, Guilin Guangxi 541004, China; 

    2. Guangxi Key Laboratory of Embedded Technology and Intelligent System, Guilin Guangxi 541004, China

  • Online:2023-04-30 Published:2023-05-01
  • Supported by:
    National Natural Science Foundation of China (62166012) 

摘要:

结直肠癌作为最常见的疾病之一,精准的结直肠息肉分割可辅助医师对其进行早期预防。然而, 在分割过程中,结直肠息肉图像存在对比度较低、病灶形状不一、位置随机化等问题,而且 Unet 网络参数量较 大但分割精度不高。因此,提出了一种基于动态平衡多尺度特征融合的 Unet 改进算法,以 Unet 为主体,结合空 洞空间卷积池化金字塔模块(ASPP)提高 Unet 深层次特征的多样性;提出通道打乱多尺度特征融合模块(CSI)和分 组多尺度特征融合模块(GI)对编解码器的卷积块进行改进,降低整体网络参数量同时提高模型的表征能力,并提 出残差金字塔拆分注意力模块(RPSA)用于编解码器的跳跃连接,平衡跳跃连接中的通道信息,提高整体网络的分 割性能。实验结果表明,该方法不仅在分割效果上优于其他方法,还大幅减少了参数量,证明了其有效性。

关键词: 动态平衡, 多尺度特征融合, Unet, 结直肠息肉, 分割算法

Abstract:

Colorectal cancer is one of the most prevalent diseases, and accurate colorectal polyp segmentation can aid physicians in early prevention. However, in the process of segmentation, colorectal polyp images present several challenges, such as low contrast, varied shapes of lesions, and randomized location. Moreover, with large parameters, the Unet network does not yield high segmentation accuracy. Therefore, an improved Unet algorithm based on dynamic balanced multiscale feature fusion was proposed. This algorithm took Unet as the main body and combined the atrous spatial pyramid pooling module (ASPP). A channel shuffle inception (CSI) module and a group inception (GI) module were also put forth to improve the convolution block of the codec, reduce the amount of network parameters, and improve the model′s characterization ability. Additionally, a residual pyramid split attention module (RPSA) was presented for the skip connection of the codec, balancing the channel information in the skip connection, and improving the overall network split performance. Experimental results showed that this method could not only outperform other methods in terms of segmentation effect, but also significantly reduce the number of parameters, thereby proving its effectiveness. 

Key words:

中图分类号: