Cell Segmentation and Morphological Analysis in Tumoroids Without Ground Truth Annotations

The morphological characteristics of cells can be studied by segmenting individual cells within a given tissue sample. Analyzing the spatial distribution and organization of cellular morphological features is crucial for understanding the architecture and functional state of the tissue. Segmenting cells in tumoroids, which are 3D culturesthat emulate a tumor’s architecture, allows for detailed morphological analysis. This is often achieved using imaging techniques based on machine learning, particularly Convolutional Neural Networks (CNNs) like U-Nets, which are suited for biomedical segmentation tasks due to their ability to capture complex spatial features. While training these algorithms typically relies on manually annotated datasets, which is a manual and effort-heavy process, it has been shown that similar performance can be achieved by training on synthetically generated data derived from a single manual segmentation. This study presents a morphological analysis pipeline for...Show moreThe morphological characteristics of cells can be studied by segmenting individual cells within a given tissue sample. Analyzing the spatial distribution and organization of cellular morphological features is crucial for understanding the architecture and functional state of the tissue. Segmenting cells in tumoroids, which are 3D culturesthat emulate a tumor’s architecture, allows for detailed morphological analysis. This is often achieved using imaging techniques based on machine learning, particularly Convolutional Neural Networks (CNNs) like U-Nets, which are suited for biomedical segmentation tasks due to their ability to capture complex spatial features. While training these algorithms typically relies on manually annotated datasets, which is a manual and effort-heavy process, it has been shown that similar performance can be achieved by training on synthetically generated data derived from a single manual segmentation. This study presents a morphological analysis pipeline for cell segmentations derived from a neural network. Following semantic segmentation via a U-Net model, instance segmentation is applied to identify individual cells within the tumoroid. To validate this pipeline, we analyze synthetic datasets generated from ground truth segmentations and introduce a metric to assess the accuracy of the instance segmentation. This pipeline is applied to real tumoroid data. Since no manual labels are available, validation is done by comparing cell features to known biological patterns and by checking if the overall structure of the tumoroids matches what is described in the literature. By removing the need for manual annotations, this approach avoids time-consuming steps and proves effective in extracting reliable information from large datasets. Rather than relying on highly precise individual segmentations, this study emphasizes the power of statistical analysis across large datasets to reveal meaningful biological trends.Show less

Leiden University Student Repository

Documents

In Collections

Cell Segmentation and Morphological Analysis in Tumoroids Without Ground Truth Annotations