Deep learning for super-resolution localization microscopy

摘要

Super-resolution localization microscopy techniques (e.g., STORM or PALM), breaks the optics diffraction limit, making possible the observation of sub-cellular structures in vivo. However, long acquisition time is required to maintain a desired high spatial resolution. To overcome the limitation, an effective method is to increase the density of activated emitters in each frame. The high-density emitters will cause them to overlap, which makes it difficult to accurately resolve each emitter location. Although some methods have been proposed to identify the overlapped emitters, these methods are computationally intensive and parameter dependent. To address these problems, in this paper, we proposed a novel method based on convolutional neural networks (CNN) for super-resolution localization microscopy, termed as DL-SRLM. DL-SRLM is capable of learning the nonlinear mapping between a camera frame (i.e., the experimentally acquired low-resolution image) and the true locations of emitters in the corresponding image region (i.e., the recovered super-resolution image). As a result, the method provides the possibility to faster resolve the high-density emitters, without requiring the parameters. To evaluate the performance of DL-SRLM, a series of simulations with varying emitter densities, signal-to-noise ratios (SNRs), and point spread functions (PSFs) were performed. The results show that DL-SRLM can accurately resolve the locations of high-density emitters, even if when the raw measurement data contained noise or was generated by using inaccurate PSF. In addition, DL-SRLM greatly improve the computational speed (~ 15 ms/frame) compared with the current methods while avoiding the effect of the parameters on the super-resolution imaging performance.