基于残差网络的无监督角膜视频分割算法
Unsupervised corneal video segmentation algorithm based on resid-ual network
白金帅 1刘凤连 1李婧 1谭左平 2汪日伟2
作者信息
- 1. 天津理工大学计算机视觉与系统教育部重点实验室和天津智能计算及软件新技术重点实验室,天津 300384
- 2. 温州理工学院浙江省巾帼科技创新工作室,浙江温州 325035
- 折叠
摘要
基于角膜形变计算出一系列生物力学特性参数是训练早期圆锥角膜分类模型的数据基础,因此圆锥角膜轮廓分割的精确性直接影响着早期圆锥角膜分类模型的准确性.本文提出了一种基于残差网络的无监督角膜视频分割方法.通过统一的网格化采样提取一组锚点被同序列视频帧所共用,从而减小网络模型学习特征表示的计算量并且提高了计算效率.同时设计了一个正则化分支对原有的视频集进行相似性转换来解决可能存在的退化解问题.与已有的无监督视频分割任务相比,本实验模型使用了少量的训练数据,但却取得了更高的分割精度和计算效率.
Abstract
The calculation of a series of biomechanical parameters based on corneal deformation is the data foundation for training early keratoconus classification models,so the accuracy of keratoconus contour segmentation directly affects the accuracy of early keratoconus classification models.In this paper,we propose an unsupervised corneal video segmentation method based on residual networks.A set of anchor points are extracted by uniformly sampling the video frames in the same sequence,which reduces the computational complexity of the network model learning feature representation and improves computa-tional efficiency.At the same time,a regularization branch is designed to transform the original video set for similarity to solve possible degenerate solutions.Compared with existing unsupervised video segmen-tation tasks,our experimental model uses a small amount of training data but achieves higher segmenta-tion accuracy and computational efficiency.
关键词
分割算法/圆锥角膜/残差网络/锚点/退化解Key words
segmentation algorithm/keratoconus/residual network/anchor point/degenerate solution引用本文复制引用
基金项目
南开大学眼科学研究院开放基金(NKYKD202209)
温州市重大科技创新攻关项目(ZG2022011)
温州理工学院科技计划研究重点项目(KY202204)
出版年
2024
NETL
NSTL
维普
万方数据