Enhanced global gradient threshold estimation model for 3D mode. Similar to Fig. S3, the training dataset for gthres estimation model for 3D mode (Fig. S8, c and d) may have limited variation. To test and improve the applicability of the model, we introduced a large dataset from a different microscope device by a different lab with image augmentation (see Fig. S3 a; n = 153) and constructed two enhanced gthres estimation models for 3D mode. (a) Reconstruction of model gthres= μG.cbg + kσG.cbg by introducing w (see Fig. S3 c) to improve compatibility to diverse images. k is replaced into a multi-variant function to apply the impact of the dispersion level of ground noise. Therefore, we name the model gthres=σG.cbg as “σcbg-w interaction model.” Left panel displays the model with the best parameters; right panel shows the impact of m and n on model performance. (b) Performance of the σcbg-w interaction model (gthres=σG.cbg) by accuracy. The correlation of the estimated gthres and MGT-determined “ground truth” gthres is measured by Pearson coefficient (r) and P-value (p). The σcbg-w interaction model has a satisfactory performance with r = 0.779. (c) Performance of the multivariate linear model gthres= a1μG.cbg + a2σG.cbg + a3w. We constructed an alternative model by a direct multivariate linear regression combining the impact of μG.cbg, σG.cbg, and w. This model has a further improved performance with r = 0.871, for users to select according to their demand. The two models are available in the library release with the σcbg-w interaction model set as default, but they are not used for the rest of the study unless specifically announced.