An instance segmentation method for detecting chain-forming phytoplankton colonies

Abstract

Microscopy-based counting of algal individuals is a fundamental procedure required for environmental monitoring and drinking-water quality assessments. Applying AI to automate such microscopic counting offers several advantages, including: (a) reducing inter-operator and inter-organization variability in evaluation, (b) decreasing labor costs, and (c) enabling surveys at unprecedented scales and frequencies, potentially leading to new scientific discoveries. However, image analysis of algae presents multiple technical challenges. In this study, we propose a novel method to accurately count both the number and sizes of relatively large chain-forming algal colonies in scenes where colonies are intricately entangled. We particularly focus on colonies that form spiral chains, aiming to appropriately estimate the number of colonies, the number of turns in each colony, and the length of each colony. Object detection in images of intertwined spirals proved difficult not only for classical instance segmentation approaches but also for newer methods such as the Segment Anything Model. Therefore, we develop a task-specific benchmark and model implementation that integrates DINOv2 with DeepONet, a physics-informed AI model. Using Mask R-CNN as a classical baseline, we demonstrate that our proposed approach substantially improves performance on the benchmark, increasing Mask IoU from 0.33 (Mask R-CNN) to 0.75.