U-FISH: a universal deep learning approach for accurate FISH spot detection across diverse datasets

Abstract

In the rapidly advancing landscape of fluorescence in situ hybridization (FISH) technologies, there is a critical need for sophisticated yet adaptable methods for spot detection. This study introduces U-FISH, a deep learning approach that significantly improves accuracy and generalization capabilities. Our method utilizes a U-Net model to transform noisy and ambiguous FISH images into a standardized representation with consistent signal characteristics, facilitating efficient spot detection. For the training and evaluation of the U-FISH model, we have constructed a comprehensive dataset comprising over 4,000 images and more than 1.6 million manually annotated spots, sourced from both experimental and simulated environments. Our benchmarks demonstrate that U-FISH outperforms existing methods for FISH spot detection, offering improved versatility by eliminating the need for laborious manual parameter adjustments. This allows for its application across a broad spectrum of datasets and formats. Furthermore, U-FISH is designed for high scalability and is capable of processing 3D data, supporting the latest generation of file formats for large and complex datasets. To promote community adoption and ensure accessibility, we provide a user-friendly interfaces: Napari plugin, web application and command-line interface. The complete training dataset is made publicly available, laying a solid foundation for future research in this field.