Low Level Feature Extraction for Cilia Segmentation
Low Level Feature Extraction for Cilia Segmentation
复制标题
DOI：
10.25080/majora-212e5952-026
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Meekail Zain;E. Miller;Shannon Quinn;Cecilia W. Lo
中科院分区：
文献类型：
--
作者：
Meekail Zain;E. Miller;Shannon Quinn;Cecilia W. Lo
纤毛是在人体某些细胞表面发现的细胞器，有节奏地扫描以运输物质。纤毛运动功能障碍通常表明纤毛病，其破坏肺、肾和其他器官内宏观结构的功能[LWL + 18]。对纤毛运动进行表型分析是了解纤毛病的重要步骤;然而，这通常是一个专家密集型过程[QZD + 15]。自动解析纤毛记录以确定有用信息的方法将大大减少所需的专家干预量。这不仅可以提高整体吞吐量，还可以减少人为错误，并大大提高基于纤毛的见解的可访问性。这种自动化很难实现，因为用于表示纤毛的图像存在噪声、部分遮挡和潜在的异相，以及纤毛占据任何给定图像的少数。纤毛的分割缓解了这些问题，因此是实现强大管道的关键步骤。然而，众所周知，纤毛很难在大多数图像中正确分割，这给管道带来了瓶颈。因此，实验和评价纤毛图像的特征提取的替代方法提供了一个更有效的分割模型的基石。目前的实验表明，使用一种新的组合特征提取器的基础分割模型的10%的改善。
—Cilia are organelles found on the surface of some cells in the human body that sweep rhythmically to transport substances. Dysfunction of ciliary motion is often indicative of diseases known as ciliopathies, which disrupt the functionality of macroscopic structures within the lungs, kidneys and other organs [LWL + 18]. Phenotyping ciliary motion is an essential step towards understanding ciliopathies; however, this is generally an expert-intensive process [QZD + 15]. A means of automatically parsing recordings of cilia to determine useful information would greatly reduce the amount of expert intervention required. This would not only improve overall throughput, but also mitigate human error, and greatly improve the accessibility of cilia-based insights. Such automation is difficult to achieve due to the noisy, partially occluded and potentially out-of-phase imagery used to represent cilia, as well as the fact that cilia occupy a minority of any given image. Segmentation of cilia mitigates these issues, and is thus a critical step in enabling a powerful pipeline. However, cilia are notoriously difficult to properly segment in most imagery, imposing a bottleneck on the pipeline. Experimentation on and evaluation of alternative methods for feature extraction of cilia imagery hence provide the building blocks of a more potent segmentation model. Current experiments show up to a 10% improvement over base segmentation models using a novel combination of feature extractors.