Morphological Segmentation and Analysis of Caenorhabditis elegans Embryo Based on Deep Learning

基於深度學習的秀麗隱桿線蟲胚胎的形狀劃分與分析

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date16 Aug 2021

Abstract

The invariant development and transparent body of the nematode Caenorhabditis elegans enable the complete delineation of cell lineages throughout its development. Despite extensive studies on cell division, cell migration and cell fate differentiation, cell morphology has not yet been systematically characterized in any metazoan, including C. elegans. This knowledge gap substantially hampers many studies in both developmental and cell biology. A well-defined morphological atlas of C. elegans could play a critical role in deciphering several biological processes, such as cell-fate asymmetry and morphogenesis.

Although recent advances in confocal microscopy have promoted in vivo four-dimensional (4D) imaging of the C. elegans embryo throughout the embryogenesis, the large quantity of volumetric image data makes the visual identification of meaningful morphological changes cumbersome. Furthermore, a comprehensive visual perception of developmental features is difficult, precluding further functional characterization. Theoretically, each cellular region can be recognized through manual annotation. However, annotating numerous volumetric images in a slice-by-slice manner is extremely time consuming, which becomes server with the number of cells and frames increasing. For example, at least one month is required for an expert to delineate an experimental dataset containing 150 frames. By addressing the challenges in cell shape reconstruction, our work aims to facilitate morphological and functional studies of C. elegans at a cellular resolution.

Our work mainly contributes in three aspects. First, we pave the way for a deep-learning-based segmentation framework by introducing a three-dimensional (3D) cell shape training dataset. To leverage the efficiency of manual annotation, a 3D membrane morphological segmentation (3DMMS) method is proposed to extract cellular masks based on their geometrical features. On the basis of an assumption that the membrane surface is smooth, 3DMMS extracts the manifold in the initial segmentation results and corrects the outliers. Concurrently, 3DMMS does not require a training dataset. However, when the membrane becomes blurry at the late developmental stage, 3DMMS has limited ability to discriminate cellular boundaries. Therefore, a training dataset is prepared in a semi-automatic manner where two experts filter and correct errors in the segmentations of 3DMMS. Finally, 4572 cells with complete 3D annotations were introduced to fuel the supervised learning network.

Second, a pipeline CShaper was designed to automate the segmentation of fluorescently labeled membranes. In the fluorescence microscopy, the attenuation of membrane signal blurs features at cellular boundary. By learning the distance transformation, a neural network DMapNet incorporates context information around boundary pixels. The learning process guarantees that cell regions are well-separated in accordance with the distance of pixels to the cell membrane. Subsequently, the membrane mask is recognized by thresholding the distance map. A group-seeded watershed algorithm is employed to partition the distance map into different cells. The identity of each cell is assigned using AceTree and StarryNite. Compared to other prevalent methods, CShaper is the only one that can detect the cavity among cells. Out of the box, the time spent on processing 150 frames is substantially reduced from hundreds of hours to ~30 minutes. For the first time, we have applied this pipeline to quantify morphological features of C. elegans based on a large-scale 3D atlas of cell morphology dataset. Statistical results on cell surface area, cell volume and cell-cell contact are in accordance with those reported in previous studies.

Finally, a large-scale morphological atlas of the C. elegans embryo is made publicly available. This atlas contains 17 digitized embryos with each cellular region explicitly identified, through which biologists can easily collect spatio-temporal shape features statistically. To the best of our knowledge, this is the largest dataset describing the morphological changes in C. elegans at single-cell level. In addition, 4572 cell annotations in 3D are also provided to facilitate further researches on the improvement of deep learning based segmentation. This dataset can enhance the studies in the fields of developmental biology, cell biology, and biomechanics.

In this study, we performed high-throughput analysis of \textit{C. elegans} to capture the diversity of morphological features. This automatic framework can enable the biologists to decipher their image data comprehensively and focus their efforts on unravelling the laws of biology.

    Research areas

  • Caenorhabditis elegans, deep learning, morphology