| 摘 要: 应用人工智能进行结肠息肉检测对医疗辅助系统十分重要,然而目前的分割模型存在准确率低、对部分样本细节难以识别的问题。因此,提出一种基于编解码结构的息肉分割模型。该模型采用跳跃轴向注意力解决梯度问题,利用适应联系训练弥补池化中信息丢失问题,使用双通道门控将低分辨率预测图细化为高分辨率显著图。在数据集CVC-ClinicDB与Kvasir-SEG上对该模型进行验证,从mIoU(平均交并比系数)、Dice(Dice相似系数)、Precision(准确率)和Accuracy (正确率)四个指标上与同类深度学习分割算法进行对比,结果为该模型在CVCClinicDB上,mIoU 为0.903,Dice 为0.947,Precision 为0.933,Accuracy 为0.933,在Kvasir-SEG 上,mIoU 为0.763,Dice 为0.868,Precision 为0.857,Accuracy 为0.867,均优于同类深度学习分割算法,验证了该模型对样本细节部分具有更好的分割效果。 |
| 关键词: 跳跃轴向注意力;适应联系训练;双通道门控;息肉分割;深度学习 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 中科院空间主动光电技术重点实验室开放基金(2021ZDKF4);上海市科委科技创新行动计划(21S31904200,22S31903700) |
|
| Research on a Novel Encoding and Decoding Structure for Colon Polyp Segmentation Algorithm |
|
LI Jun, WANG Fang, YANG Haima, SONG Yeye
|
(School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China)
lijuny@usst.edu.cn; wovf97@outlook.com; snowyhm@sina.com; syyhut@163.com
|
| Abstract: The application of artificial intelligence for colon polyp detection is very important for medical assistance systems. However, current segmentation model has problems of low accuracy and difficulty in recognizing some sample details. Therefore, this paper proposes a polyp segmentation model based on encoding and decoding structure. In this model, skip axial attention is used to solve the gradient problem, adaptive connection training is used to compensate for information loss in pooling, and dual channel gating is used to refine the low resolution prediction map into a high-resolution saliency map. The model is verified on datasets CVC-ClinicDB and Kvasir-SEG and Compared with similar deep learning segmentation algorithms on four metrics: mIoU, Dice, Precision, and Accuracy. Results show that on CVC-ClinicDB, mIoU is 0.903, Dice is 0.947, Precision is 0.933, and Accuracy is 0.933. On Kvasir-SEG, mIoU is 0.763, Dice is 0.868, Precision is 0.857, and Accuracy is 0.867. All of the results are superior to similar deep learning segmentation algorithms, verifying that the proposed model has better segmentation performance for sample details. |
| Keywords: skip axial attention; adaptive connection training; dual channel gating; polyp segmentation; deep learning |
