Bayesian Influence and Physically Motivated Noise Models in Three Dimensional Structured Illumination Microscopy

Structured Illumination Microscopy (SIM) has been used to improve images of biological structures beyond what is possible with wide-field imaging, and this improvement can be taken across many planes to construct a 3D image. However, for a method to work across SNR regimes and truly approach the desired theoretical resolution, a physically accurate forward model (including noise model) must be used. Here, we present Bayesian Inference Optimizer-3DSIM (BIO-3DSIM), a Pytorch based parallelized optimizer for reconstruction of 3DSIM images. BIO-3DSIM accounts for both photon noise and detector noise in its optimization scheme and finds the most likely imaged structure based on knowledge of the apparatus’ Point Spread Function (PSF) and camera model. BIO-3DSIM implements full parallelization of its optimization scheme across the GPU to reduce computational expense. With BIO-3DSIM’s physical forward model approach, we improve reconstruction resolution, especially in cases of low SNR, while not significantly inflating runtime so as to maintain convenience of use.