Unified Multi‑Task Foundation Model for C. elegans

Abstract

Recent advances in deep learning have transformed biological image analysis, yet most existing models are task-specific and computationally intensive, limiting their scalability in real-world laboratory workflows. Here, we propose a unified and lightweight foundation model tailored for Caenorhabditis elegans, capable of performing multiple vision tasks including semantic and instance segmentation, object detection, keypoint and arbitrary-point tracking, brightfield and fluorescence denoising, and super-resolution, all within a shared backbone architecture. The model is trained on a comprehensive dataset integrating public and in-house imaging data, covering diverse imaging modalities and biological contexts. To evaluate biological relevance and generalization, we design three benchmark applications: drug screening, transgenic phenotyping, and behavior–neural activity coupling. Our approach emphasizes high performance, efficiency, and deployability, enabling real-time analysis in high-throughput experimental settings. This work establishes a modular and scalable foundation for multi-task visual inference in C. elegans, offering a broadly applicable framework for lightweight foundation models in biological imaging.

Keywords