Automated spheroid and organoid image analysis from rule-based methods to deep learning
- 1Department of Pharmacology, School of Medicine, Faculty of Health Sciences, University of Pretoria, Pretoria, 0007, South Africa; Corresponding author.
- 2Department of Physiology, School of Medicine, Faculty of Health Sciences, University of Pretoria, Pretoria, 0007, South Africa
- 3Department of Pharmacology, School of Medicine, Faculty of Health Sciences, University of Pretoria, Pretoria, 0007, South Africa; African Institute of Data Science and Artificial Intelligence, University of Pretoria, Pretoria, South Africa
Abstract
The increased adoption of three-dimensional (3D) cultures in high-content screening has created an analytical bottleneck in image analysis, prompting the development of automated image analysis tools. This review examines the evolution of automated tools for analysing brightfield and phase-contrast images of spheroids and organoids, from semi-automated approaches through classical machine learning pipelines to modern deep learning tools. Each generation is evaluated according to its underlying principles and limitations, especially when applied to heterogeneous datasets. Rule-based tools were introduced to automate the measurement of morphometric parameters established in manual workflows but may not generalise when experimental conditions change. Classical machine learning improved robustness to post-treatment morphologies but requires retraining when imaging conditions change. A range of deep learning architectures that learn directly from datasets has been introduced, partially reducing the retraining burden associated with classical machine learning. No generation supersedes another across all experimental contexts, and tool selection should be guided by the biological endpoint, image conditions, culture morphology, annotation availability and computational resources. Most tools analyse two-dimensional projections of inherently 3D structures, reducing the accuracy of volume estimates for irregular structures. Independent benchmarks demonstrate that tool performance varies across datasets and imaging laboratories. Future tool development should prioritise robustness across imaging conditions, identification of segmentation failure modes and validation on shared datasets.
Related
- SpheroQuaNNt: U-Net–based volume estimation in tumor spheroid and organoid models for assessing preclinical treatment response
- Automated phytoplankton single‐cell counting from digital images with transfer learning for harmful algal bloom monitoring in the English Channel and North Sea
- Combinatorial Network-Based Manifold Topological Deep Learning for Image Analysis
- Automated Platelet Classification Under Different Solvent Conditions Through Deep Learning.
- Evolution of Vesicle Shape Parameters Based on A Deep-Learning Based Image Analysis Framework Over Extended Time Periods
BibTeX
@article{ncube2026automated,
title = {{Automated spheroid and organoid image analysis from rule-based methods to deep learning}},
author = {Keith Ntokozo Ncube and Iman van den Bout and Werner Cordier},
journal = {SLAS Discovery},
year = {2026},
month = oct,
doi = {10.1016/j.slasd.2026.100344},
url = {https://doi.org/10.1016/j.slasd.2026.100344},
note = {CC BY},
}
RIS
TY - JOUR
TI - Automated spheroid and organoid image analysis from rule-based methods to deep learning
AU - Keith Ntokozo Ncube
AU - Iman van den Bout
AU - Werner Cordier
PY - 2026
DA - 2026/10/30
JO - SLAS Discovery
DO - 10.1016/j.slasd.2026.100344
UR - https://doi.org/10.1016/j.slasd.2026.100344
AB - The increased adoption of three-dimensional (3D) cultures in high-content screening has created an analytical bottleneck in image analysis, prompting the development of automated image analysis tools. This review examines the evolution of automated tools for analysing brightfield and phase-contrast images of spheroids and organoids, from semi-automated approaches through classical machine learning pipelines to modern deep learning tools. Each generation is evaluated according to its underlying principles and limitations, especially when applied to heterogeneous datasets. Rule-based tools were introduced to automate the measurement of morphometric parameters established in manual workflows but may not generalise when experimental conditions change. Classical machine learning improved robustness to post-treatment morphologies but requires retraining when imaging conditions change. A range of deep learning architectures that learn directly from datasets has been introduced, partially reducing the retraining burden associated with classical machine learning. No generation supersedes another across all experimental contexts, and tool selection should be guided by the biological endpoint, image conditions, culture morphology, annotation availability and computational resources. Most tools analyse two-dimensional projections of inherently 3D structures, reducing the accuracy of volume estimates for irregular structures. Independent benchmarks demonstrate that tool performance varies across datasets and imaging laboratories. Future tool development should prioritise robustness across imaging conditions, identification of segmentation failure modes and validation on shared datasets.
SN - 2472-5552
SN - 2472-5560
ER -