Phytoplankton are microbes driving half of global primary production and are widely used model organisms in various studies. Despite advancements in high-throughput methods for acquisition of cell size and abundance data, microscopy remains widely preferred in multi-strain experiments. In this study, an instance segmentation model was trained on a dataset containing 11 strains and integrated into an image processing pipeline. The final version of the pipeline showed superiority over an established thresholding method (Avg. F1 Score 86% vs. 34%). This approach also highlights the potential of clustering unseen strains by prompting a general-purpose segmenter with low-confidence predictions.
This work is licensed under a Creative Commons Attribution 4.0 International License.