Figure 7.

PatchWarp algorithm for fully automatic across-session registration

(A) Registration between different imaging sessions for longitudinal analyses (red: session G images; cyan: session H images). Mean- and max-projection images were used as the inputs to the algorithm. The first step finds the optimal Euclidean transformation that corrects large displacements with rigidity (translation and rotation). The second step finds the optimal affine transformations for small patches to correct non-uniform distortions between the two image sessions. The transformations are separately estimated for mean and max-projection images, but only one transformation with larger ECC between G and transformed H, $\left({\rho }_{G,E\left(H\right)},{\rho }_{G_i,A_i\left\{E\left(H_i\right)\right\}}\right)$, is selected at each step. The obtained transformations can be used to transform any images from session H to the session G coordinate for longitudinal neural activity analyses.

(B) Zoomed images of the subfield that is indicated by yellow dashed lines in (A). This subfield highlights the necessity of the warp-correction step.

(C) Correlations of mean or max-projection images between early and late imaging sessions of the same neural population (six different FOVs, cell body imaging, glass window). The correlation is compared between pre-registration (raw), post-Euclidean transformation (rigid), and post-patchwork affine transformations (warp). Warp correction consistently improved the registration accuracy between different imaging sessions. Each thin gray line indicates the comparison of the same imaging session across conditions. The error bars are 95% CI.