| 摘 要: 针对结直肠息肉图像中由于病灶形状不规则导致分割边界模糊及分割精度低的问题,提出了一种基于混合注意力和金字塔池化的结直肠息肉分割算法PC-TransUnet(Pyramid Convolutional TransUnet)。首先,采用金字塔池化模块(Pyramid Pooling Module,PPM)捕获不同尺度的上下文信息,从而增强网络的感受野;其次,利用混合注意力模块(Convolutional Block Attention Module,CBAM)关注重要的语义特征并抑制无关的信息;最后,将解码器中的激活函数ReLU替换为SiLU,通过增加激活范围进一步提高分割精度。在Kvasir-SEG和CVC-Clinic DB数据集上的实验结果显示,该算法的平均相似性系数分别达到了96.99%和96.41%,较基础网络分别提高了2.91百分点和2.96百分点。经过实验验证,PC-TransUnet在提高分割精度、减少分割边界模糊方面表现出色,整体性能均优于当前的主流算法。 |
| 关键词: 结直肠息肉分割;混合注意力;金字塔池化;SiLU |
|
中图分类号: TP394
文献标识码: A
|
| 基金项目: 国家自然科学基金(61971272,61601271);国家重点研发重点专项(2019YFC1520204);国家自然科学基金青年科学基金项目(61601271);陕西省教育厅专项科研计划项目(15JK1086) |
|
| Colorectal Polyp Segmentation Algorithm Based on Improved TransUNet |
|
WANG Yuhang, YANG Yun
|
(College of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, X'i an 710021, China)
1339139550@qq.com; yangyun11@163.com
|
| Abstract: To address the issues of blurred segmentation boundaries and low accuracy caused by irregular lesion shapes in colorectal polyp images, this paper proposes a colorectal polyp segmentation algorithm PC-TransUNet (Pyramid Convolutional TransUnet) based on Hybrid Attention and Pyramid Pooling. Firstly, a Pyramid Pooling Module (PPM) is employed to capture multi-scale contextual information, thereby enhancing the receptive field of the network. Secondly, a Convolutional Block Attention Module (CBAM) is utilized to focus on critical semantic features while suppressing irrelevant information. Finally, the ReLU activation function in the decoder is replaced with SiLU to improve segmentation accuracy by expanding the activation range. Experimental results on the Kvasir-SEG and CVCClinicDB datasets demonstrate that the proposed algorithm achieves average Dice coefficients of 96.99% and 96.41%,respectively, representing improvements of 2.91 percentage points and 2.96 percentage points over the baseline network.Extensive validation confirms that PC-TransUNet significantly enhances segmentation accuracy, mitigates boundary blurring, and outperforms existing mainstream algorithms in overall performance. |
| Keywords: colorectal polyp segmentation; Hybrid Attention; Pyramid Pooling; SiLU |
