Self-contrastive learning enables interference-resilient and generalizable fluorescence microscopy signal detection without interference modeling

Abstract

Every weak signal in fluorescence microscopy may contain critical biological information. However, the inter-ference resilience required to detect such signals has traditionally relied on task-specific interference modeling, which limits generalizability. Here, we present a self-contrastive learning-based signal detection solution that achieves interference resilience without the need for interference modeling, thereby offering high generalizability. The method, DEep PAttern Fitting (DEPAF), is a module that contrasts asynchronously generated data views from asymmetric model paths to extract signals from interference, while incorporating highly parallel signal recognition and localization in the process. Benchmarking shows that DEPAF substantially improves signal detection performance across diverse imaging modalities and dimensions, especially under challenging conditions such as low SNRs and ultra-high signal densities. It is also compatible with, and consistently enhances the performance of various imaging techniques, such as super-resolution imaging, spatial transcriptomic imaging, and two-photon calcium imaging. Notably, DEPAF relies only on image patches with signal patterns and one tunable hyperparameter to adapt to new tasks, making it accessible to users without domain-specific expertise and lowering the barrier to broader adoption. DEPAF is expected to advance the signal-centric fluorescence microscopy techniques and inspire further advancements, especially in the era of image-based multi-omics.