Workflow from Scientific Research

Upon the acquisition of camera frames, detected single-molecule emission patterns from stochastic lateral and axial positions are isolated and sent to a trained DNN. The network outputs a vector of mirror deformation-mode amplitudes, for each detection of a single molecule. The estimations before and after each compensation are then combined through a Kalman filter to drive the next deformable mirror update. p and q represent numbers of feature maps input and output to a residue block (the orange box). N represents the image width/height. s is stride size in a convolutional layer. The detailed sizes in each layer of the network architecture can be found in Supplementary Table 1 .

DOI