Application of Unsupervised Clustering Techniques to Somatic Embryos for Automated SE Fluidics System
Year: 2026
Type: preprint
Abstract: Somatic embryogenesis (SE) is a viable method for clonal propagation of economically significant crops including forest trees such as pines and spruces. However, large-scale SE production requires automation to reduce manual labor, especially in selecting and harvesting mature embryos. Since embryo maturation is asynchronous, identifying viable embryos remains a challenge. This study explores machine learning techniques to cluster grayscale images of mature SEs processed through a fluidics-based SE system for image analysis and sorting.A dataset of 720 grayscale embryo images with a subset of embryos labelled as 'accepted' or 'rejected' based on the embryo morphology was selected for this study. The full dataset was clustered using a range of UL methods. The K-means method, which requires human intervention to select the k value, with Principal Component Analysis (PCA) achieved 82% accuracy for rejected and 92% for acceptable embryos. Constrained Unsupervised Learning (CUL), where pairwise constraints were introduced to nudge the UL method to improve the clustering accuracy of segregating clusters, compared to morphology-based clustering achieved nearly 100% accuracy for acceptable embryos and 87.5% for rejected embryos. Using an unconstrained hierarchical clustering technique, dendrograms were plotted and analyzed to understand cluster formations at different threshold values for new image datasets that consisted of different numbers of grayscale images (500, 3,500, and 4,089). It was found that the dataset consisting of 500 images sufficiently represents the exhaustive morphological traits of the embryos. The results show that a small number of embryos in the range of few hundred can be used to provide the base cluster that can be used for identifying embryos with acceptable morphology. Lastly, a supervised learning (SL) model was trained on the clusters extracted from CUL, achieving 100% segregation of rejected embryos. The trained SL model was32% faster than morphological methods. The proposed pipeline ensures efficient in-situ embryo categorization, facilitating scalable automation in industrial SE applications while minimizing human bias.
Source: Research Square
Institution: Georgia Institute of Technology
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7166198837