Automated tracking of the brown algal parasite Eurychasma dicksonii with deep learning

This study presents an automated real-time tracking approach for Eurychasma dicksonii, an intracellular parasite affecting brown algae, using deep learning applied to time-lapse microscopy video analysis. Traditionally, parasite detection and tracking rely on manual observation, resulting in a time-consuming, operator-dependent process prone to inaccuracies. To overcome these limitations, we developed and evaluated a comparative deep learning framework for parasite segmentation and multi-object tracking in microscopy videos. Specifically, we assessed five segmentation architectures, including convolution-based YOLO models (YOLOv8 and YOLOv11), a YOLO-transformer-based model (YOLOv8-SwinT), and two RF-DETR-based models (RF-DETR-Seg nano and RF-DETR-Seg large), to examine how different detection paradigms perform under challenging microscopy conditions. The best-performing segmentation model was then integrated with multi-object tracking algorithms to detect and follow parasite cells across video frames. By combining segmentation with tracking, the system better handles occlusions, noise, and morphological variability, improving robustness and precision in parasite detection and motion analysis. Experimental results show that RF-DETR-Seg large achieved the best segmentation performance, reaching an mAP50 of 74.0% for boundary and 70.9% for bounding box detection, and achieving the highest precision among all evaluated models. When integrated with tracking algorithms, the RF-DETR-Seg large + ByteTrack combination achieved the most robust performance (MOTA = 66.08%, IDF1 = 75.47%), indicating strong tracking accuracy and identity consistency. To demonstrate biological utility, we implemented a downstream feature extraction module to quantify parasite expansion dynamics from segmentation mask area time series. Applied to four parasites across two representative test videos, the pipeline enabled quantitative characterization of parasite behavior at the single-object level. Overall, this study highlights the potential of AI-driven methods to accelerate parasitology research by providing quantitative insights into parasite dynamics and a reproducible framework for microscopic imaging and biological analysis.