Fig. 3
From: A large-scale curated and filterable dataset for cryo-EM foundation model pre-training
Details of Motion Correction Pipeline. (a) The pipeline starts with the input of raw movies and their corresponding gain reference. First, gain correction is applied to address the detector’s non-uniform response. Then, We use the patch motion correction algorithm in CryoSPARC to estimate motion for each frame, followed by motion correction and alignment of even and odd frames separately. Notably, CryoSPARC performs background estimation and background subtraction after the motion correction. Finally, to reduce data storage overhead, the even micrograph is subtracted from the odd micrograph to generate the full-diff MRC pair. (b) The left image shows the patch-wise motion estimation for all frames, with the starting frame in blue and the ending frame in yellow. The global rigid motion and local patch-level bending motion are combined and amplified by a factor of 20 for visualization. The right image displays the pixel-wise optical flow estimation obtained through spline interpolation, as well as the patch-wise motion direction of the current frame relative to the previous frame. (c) We generate the full-even-odd triplet by combining motion-corrected even and odd frames. This approach supports Noise2Noise training. (d) We perform background estimation on the motion-corrected images to separate and subtract the background, enhancing the image contrast.
