Improving single molecule localisation microscopy reconstruction by extending the temporal context

Single-molecule localization microscopy methods such as d STORM require specific buffer conditions to enable blinking and detection of individual emitters, making them incompatible with live cell imaging and expansion microscopy. An alternative approach to achieve super-resolution without blinking is to observe the fluctuations of the emitter intensity over time. Existing localization algorithms for high-emitter density make use of radial symmetry or use artificial neural networks trained on single high-density frames to predict emitter positions. Here, we aim to improve the resolution by using a larger temporal context. We combine the U-Net architecture used previously for image reconstruction with multi-head attention used in the Transformer architecture. We compare the results to DECODE and eSRRF as well as to traditional fitting algorithms on public benchmark data. A generic pre-trained model is provided together with a fast and robust simulator for training data and all scripts needed to train custom networks.