Quantitative Characterization of Individual Algae Cells via Machine Learning based Clustering

The industrial adoption of microalgae cultivation is currently limited by economic feasibility and the challenge of accurately predicting growth behavior. Conventional control philosophies often focus on isolated parameters rather than a holistic, predictive assessment of culture status. This study addresses this gap by extracting morphological information at the single-cell level to characterize population dynamics. We used a pipeline for cell segmentation and feature extraction, capturing diverse properties including shape, color, texture, and intensity distribution. To evaluate the underlying structure of this data, three distinct clustering techniques were compared: density-based (DBSCAN), distance-based (K-Means), and distribution-based (GMM). Our findings indicate that DBSCAN is not suited for this application, as the continuous nature of microalgal growth prevents the formation of discrete density-based clusters. In contrast, both K-Means and GMM successfully processed the data. By setting a cluster size of eight, we identified three highly consistent clusters across both algorithms, corresponding to distinct, identifiable phenotypes validated by domain knowledge. However, the remaining five clusters showed significant mismatches, highlighting the sensitivity of phenotype classification to the underlying mathematical assumptions. These results demonstrate that while machine learning offers a powerful tool for holistic culture monitoring, the choice of algorithm is critical for interpreting biological states.

Keywords: Microalgae, Image Analysis, Automated Phenotyping, Cell Segmentation, Growth Prediction.