EvoAgents: Generative AI for Rapid Architecture Search in Cell Nuclei Detection

Science relies on image data across several disciplines like microbiology, but developing effective deep learning models for segmentation and detection tasks present challenges, including limited time, expertise, or financial resources. We introduce an evolutionary agent that autonomously generates, evaluates, and evolves convolutional neural networks (CNNs) based on natural language descriptions. Powered by large language models (LLMs) and guided by a genetic algorithm, our agent allows scientists to obtain trainable, reproducible model code for their specific tasks without manual development. A case study on cell nuclei detection demonstrates the pipeline’s feasibility and therefore establishes a baseline for architectures driven by LLMs. Designed for generalizability to diverse scientific imaging problems, this agent acts as a collaborative interface between deep learning and science and therefore contributing to scientific discovery and reduced economic and technical barriers.