An illustrated introduction of the actin image simulation model to generate 3D cytoskeleton images with ground truth. (a) A schematic diagram describing the image simulation model. The training dataset is the same as Fig. 2. As preparation, raw images were processed by 3D ILEE. To mostly simulate the shape of actin, we directly adapted the skeletonized images (I_sk) generated from 3D ILEE as the central line of new filaments but filled all voxels with novel values. The artificial images comprise three brightness fractions: artificial actin, artificial ground noise, and artificial diffraction noise. First, referring to the binary images of the training samples, for each voxel (I) in the filament, we found its referred voxel (I_R), the nearest voxel to I in the filament skeleton. (b and c) Next, we calculated the D_R (the Euclidian distance between I and I_R) and D_R2N (the Euclidian distance from IR to the nearest negative voxel) of all voxels that belongs to the filament (also see b) and constructed a polynomial regression model (also see c) between the relative distance to I_R (i.e., D_R/D_R2N) and the mean (μ_atf) as well as the STD(σ_atf) of relative brightness (i.e., I/I_R) for all on-filament voxels for all images. To generate the new artificial actin fraction, we refilled the on-filament voxel by the normal distribution I/IR ∼ N(μ_atf,σ_atf²) based on the brightness of voxels belonging to the corresponding I_sk, and we abandoned the voxels with D_R/D_R2N >1.4 because they are more likely to be the structural noise—the cytosol and PM with fluorescence markers not binding to the actin. Second, for the ground noise fraction, we had different strategies to generate two groups of voxels. For the volume previously belonging to the coarse background, we directly apply a fold of change that is subject to a normal distribution to each voxel; for the other volume beyond coarse background, we gave random values that subject to a Burr-12 distribution fitted by voxels of the coarse background, thereby all voxels in the new artificial image were given with a ground noise value. Third, the volume previous belonging to neither coarse background nor real actin is considered as the space impacted by the diffraction (Fig. 2 a). Using the raw data of this region, we fitted an exponential distribution to describe the diffraction signal and refilled the diffraction region by this distribution. The choice of Burr-12 and exponential distribution is based on the best performance over a fitting test over 60 statistical models. Finally, we add the three fractions of brightness together and linearly align the value range to 0-4095 (12 bit). (b) a schematic diagram explaining the data utilized to fit the actin simulation model. (d) (c and d) the fitting result of the expected Mean of STD of the relative brightness of the actin simulation model. Only data with D_R/D_R2N lower than 1 were used to fit the model because they were less likely to be influenced by the structural noise. As D_R/D_R2N were discretely distributed, we added weights equal to log₁₀N, where N means the number of voxels among the 31 samples, to each D_R/D_R2N.