Enhanced 3D NNES-based adaptive coarse thresholding model for samples with greater diversity. As the original training dataset of Fig. S2 may not have satisfactory variation of threshold-at-peak range, we introduced published actin image samples (Cao et al., 2022) from a different microscope device in a different lab and applied image augmentation to generate a greatly diverse dataset for model training, resulting an enhanced 3D coarse thresholding model. (a) Scheme of datasets. 31 and 20 random samples, by different microscope machines and photographers, are collected. The total 51 images are then processed through a linear LUT transformation (see formula) to repress (k1 = 0.7) or enhance (k1 = 1.3) the ground noise (see altered brightness distribution, right panel). (b) the correlation between global threshold at NNE peak (p) and coarse global threshold (tcbg) measure through MGT, of the total training dataset. The black line suggests the “1st edition” model (Fig. S2) used for the rest of this study, which fits very well with all raw and augmented samples from our lab collection but not with the other dataset. (c) Differences in brightness distribution between datasets. We inspected both datasets to investigate why they have different slope (see b). It is determined that our lab collection have a single-pattern, condensed distribution of ground noise brightness, while the Cao dataset has a wide-spread, dispersive peak. This is because the Cao dataset has multi-layer ground noises, as some of the cells have higher baseline brightness potentially due to out-of-focus light or non-binding actin fluorescence marker. Arrows with different color indicates two background areas with different ground noise level. (d) Performance of enhanced 3D coarse global threshold model . As the ground noise concentration level have an impact to the relationship between p and tcbg, we measured a second feature value—the distance between two points with 80% top frequency (w; see c)—for modeling improvement. To introduce of w, we constructed a new model , where kcbg, m, and n are trained constants. The left panel shows the correlation between and tcbg at the best performance (); the right panel shows the relationship between m, n, and correlation coefficient R2. Note: This model will be provided as the default model for 3D tcbg estimation in the PyPI library release but not used for the rest of the study unless specifically announced.