aiSEGcell: User-friendly deep learning-based segmentation of nuclei in transmitted light images

Conventional cell segmentation approaches use fluorescence to mark cells and cell compartments for segmentation (Fig 1a). In contrast, aiSEGcell omits fluorescence markers for segmentation and directly segments BF images (Fig 1b). We selected U-Net as the model architecture [12,14] for aiSEGcell and trained it on 9,849 BF image and segmentation mask pairs (>600,000 nuclei) of live primary murine hematopoietic cells, like macrophages, stem and progenitor cells, at 20x magnification (D1) [16]. To highlight the need for user-friendly software that segments nuclei in BF, we compared aiSEGcell to widely used nucleus segmentation software Cellpose and StarDist [19,20]. Available pretrained models for nucleus and BF segmentations were not suitable alternatives to the specialized BF nucleus segmentation of aiSEGcell (S1 Fig and S1 and S2 Tables). We evaluated aiSEGcell on a second data set (D2) comprised of 6,243 image-mask pairs (>300,000 nuclei) derived from five independent imaging experiments with experimental settings similar to D1. Cell morphologies included round cells and more complex shaped cells with protruding pseudopodia. In addition, cell sizes varied by an order of magnitude (Table 1). aiSEGcell segmented nuclei in BF images across morphologically distinct cells and different scales (Figs 1c and S2a). We quantified segmentation quality using an adapted version of the F1-score that counted inaccurately predicted objects (e.g. intersection over union = 0.4) as one error instead of as both, false positive (FP), and as false negative (FN; see Materials and methods). Therefore, the image-wise average adapted F1-score with an intersection over union (IOU) threshold of 0.6, denoted as , more appropriately represented the observed segmentation quality [21]. aiSEGcell nuclear segmentations in BF were quantitatively similar to nuclear marker derived segmentations (ground truth; = 0.66). Around 70% versus 5% of images had F1^0.6≥0.6 versus F1^0.6≤0.4, illustrating that aiSEGcell segmentations are consistently accurate across challenging typical BF images (Fig 1d and S3 Table). Furthermore, we quantified the frequency of oversegmentations (an object like a nucleus in the ground truth segmentation is split into multiple objects in the aiSEGcell segmentation) and undersegmentations (multiple objects in the ground truth segmentation are merged into a single object in the aiSEGcell segmentation). Only 0.2 (s.d. = 0.5) of 50 ground truth nuclei were oversegmented in aiSEGcell segmentations and 0.2 (s.d. = 0.6) of 50 nuclei detected in aiSEGcell were merged ground truth nuclei per image on average in D2 (S4 Table). Images in D2 were intended for cell tracking and consequently not of high cell density. To evaluate if undersegmentations were more frequent in images dense with cells, we acquired an additional data set (D7) with densely seeded cells. aiSEGcell accurately ( = 0.80) segmented nuclei in BF and undersegmentations rarely occurred. Only 0.2 (s.d. = 0.5) out of 50 nuclei detected by aiSEGcell were merged ground truth nuclei per image on average (S3 Fig and S4 and S5 Tables).

Fig 1. aiSEGcell accurately segments nuclei in bright field images.

(a) Workflow to segment nuclei based on nuclear marker or (b) BF (cyan; scale bars 40 μm). (c) Nuclear segmentations in BF were qualitatively similar to fluorescent nuclear marker derived segmentations. Each field-of-view stems from a different independent experiment of data set D2 (scale bars 10 μm). (d) Nuclear segmentations in BF (b) were quantitatively similar to nuclear marker derived segmentations (a) in D2 (n = 6,243 images, N = 5 experiments). The F1-score ranges between 1 (best) and 0 (worst) and the image-wise average F1 with an intersection over union threshold of 0.6 is denoted as . Given the challenges of imaging live primary non-adherent cells we consider F1^0.6 ≥0.6 good. The results in Figs 2, 3c and 4c support this estimate. Images were cropped from a larger field-of-view and contrast was individually adjusted here for visibility.

https://doi.org/10.1371/journal.pcbi.1012361.g001

Mask-level performance evaluation of segmentation models may not be relevant to illustrate suitability for downstream applications. Consequently, evaluating segmentation models with respect to biologically meaningful read-outs is necessary. The translocation of fluorescently labeled signaling molecules from the cytoplasm to the nucleus and vice versa, in small motile live primary cells imaged with low signal-to-noise is demanding, sensitive to variations in the nuclear mask and has been shown to affect cell behavior and fate [22–24]. We used a homozygous reporter mouse-line that had all p65 proteins, the transcriptionally active part of the nuclear factor κB (NfκB) dimer, tagged with green fluorescent protein (GFP) to quantify its translocation to the nucleus upon NfκB signaling activation (Fig 2a). We imaged highly motile live primary macrophages, hematopoietic stem and progenitor cells in conditions optimized for cell health and acquisition speed resulting in low signal-to-noise images—a realistic challenging experimental setting that tests the feasibility of aiSEGcell segmentations for many other applications. Previous research showed that hematopoietic cells stimulated with tumor necrosis factor α (TNFα) respond with either of four characteristic NfκB signaling dynamics response types: non-responsive (NON), oscillatory (OSC), sustained (SUS), or transient (TRA; Fig 2b) [25]. We tracked 458 single cells exposed to control media or TNFα for up to 23 h in the five independent time lapse experiments of D2 (Table 1). With processing and quantification otherwise identical, we compared the NfκB responses derived from aiSEGcell (trained on D1) and ground truth segmentations. We observed that single cell signaling dynamics classifications based on aiSEGcell segmentations closely matched those based on ground truth segmentations (average Euclidean distance = 0.53, n = 458; Fig 2b). Moreover, 81% of aiSEGcell signaling dynamics were assigned the same response type as the corresponding signaling dynamics obtained from ground truth segmentations. Classifications were accurate for all response types (NON: 88%, OSC: 77%, SUS: 79%, TRA: 74% match) and for the different experiments contributing to D2 (Figs 2c and S4). Consequently, aiSEGcell segmentations do not introduce a bias towards NfκB response types and retain the characteristic response type frequencies of different primary hematopoietic cells.

Fig 2. aiSEGcell is sufficient to retain demanding biologically meaningful information.

(a) NfκB signaling dynamics obtained from a representative single tracked primary murine granulocyte/monocyte progenitor cell. The magenta bar and dashed line illustrate a single addition of TNFα after 1 h. Contrast was individually adjusted here to improve visibility (scale bars 10 μm). (b) Signaling dynamics in D2 divide into four response types: non-responsive (NON), oscillatory (OSC), sustained (SUS), and transient (TRA). Signaling dynamics obtained from D1-trained aiSEGcell (cyan) were qualitatively similar to manually curated nuclear marker derived signaling dynamics (black) [25]. Exemplary signaling dynamics from all independent experiments of D2 are displayed with ground truth response types. The length of signaling dynamics varies due to different experimental settings in D2 experiments. Euclidean distance (D) illustrates similarity between ground truth and aiSEGcell signaling dynamics. Larger D correspond to larger dissimilarity between signaling dynamics and D = 0 for identical signaling dynamics. (c) Single cell signaling dynamics response type assignments were similar between aiSEGcell-derived and fluorescent nuclear marker-derived signaling dynamics. Four aiSEGcell signaling dynamics were classified as outliers due to many missing segmentation masks and were omitted from the confusion matrix for clarity (n = 458 signaling dynamics, N = 5 experiments).

https://doi.org/10.1371/journal.pcbi.1012361.g002

Once trained, neural networks have the potential to be applied indefinitely to data of the same distribution, in some cases even to new data distributions. This training process is notoriously challenging and requires hardware resources that are prohibitive for many users. Consequently, it is important to investigate the robustness of a neural network to unseen data. We tested aiSEGcell trained on D1 on a published unseen data set (D5) comprised of human multipotent progenitor populations (hMPP) cultured in a microfluidic chip acquired by an unseen experimenter on an unseen microscopy setup (Figs 3a and S2b and Tables 1 and S6). We repeated the published data processing to obtain single cell extracellular signal-regulated kinase (ERK) pathway signaling dynamics, now using aiSEGcell segmentation masks. Finally, we automatically assigned ERK signaling dynamics to one of four response types: NON, TRA, intermediate (INT), SUS [22,26]. ERK signaling dynamics obtained from aiSEGcell segmentations were similar to those based on ground truth segmentations (average Euclidean distance = 0.50, n = 126; Fig 3b). Overall, the signaling dynamics classifications corresponded very well between manually curated or aiSEGcell segmented nuclei, illustrating aiSEGcell’s usability: 81% of aiSEGcell response type assignments matched the ground truth signaling dynamics (NON: 94%, TRA: 36%, INT: 69%, SUS: 88%; Fig 3c). TRA signaling dynamics had a lower classification overlap. However, closer inspection of their dynamics data showed that ‘misclassified’ aiSEGcell signaling dynamics curves were actually very similar to those from manually curated nucleus segmentations. They had a good average Euclidean distance of 0.51 demonstrating the similarity of aiSEGcell and ground truth signaling dynamics. Thus, the differences in classifying a TRA response do not stem from clear differences in manually curated versus aiSEGcell segmentation, but from the high sensitivity of the used classification parameters to minor changes in curve properties (due to the used cut-offs of signal returning to baseline) [26]. As shown in S5 Fig, these very small segmentation deviations would likely not be considered biologically relevant by an expert.

Fig 3. aiSEGcell accurately segments in unseen experimental conditions and is retrained with little data.

(a) Exemplary segmentations of hMPPs. D1-trained aiSEGcell segmentations (cyan) were qualitatively similar to nuclear marker derived segmentations (black) on unseen experimental conditions. Human MPPs were imaged on a microfluidic chip (scale bar 10 μm). (b) ERK signaling dynamics in hMPPs divide into four response types: non-responsive (NON), transient (TRA), intermediate (INT), and sustained (SUS). Euclidean distance (D) illustrates similarity between manually curated ground truth and aiSEGcell signaling dynamics. (c) Signaling dynamics response type classifications were similar between aiSEGcell and ground truth signaling dynamics (n = 126, N = 1 biological replicate). (d) aiSEGcell trained on D1 was retrained on D3 (n = 34 images, 4,083 cells total, 237 cells with nucleus mask>750 μm²) and D4 (n = 32 images, 3,815 cells) to segment mMEGs (green) and mESCs (purple), respectively. (e-f) Retraining aiSEGcell qualitatively (e) and quantitatively (f) improved nuclear segmentations of mMEGs in BF (scale bar 40 μm). IOU was only computed for large cells (nucleus mask>750 μm²; n = 21 unique field of views, N = 1 biological replicate; two-sided paired Wilcoxon signed-rank test, W = 0, p = 9.5e-7; ***: p<0.001). In boxplots, the central black line is median, box boundaries are upper and lower quartiles, whiskers are 1.5x interquartile range. (g-h) Retraining aiSEGcell qualitatively (g) and quantitatively (h) improved nuclear segmentations of mESCs in BF (scale bar 30 μm; n = 10 unique field of views, N = 1 biological replicate; two-sided paired t-test, degrees of freedom = 9, t = -124.4, p = 7.1e-16). Images were cropped from a larger field-of-view and contrast was individually adjusted here to improve visibility.

https://doi.org/10.1371/journal.pcbi.1012361.g003

Small focus changes due to technical problems or cell differentiation are usually detrimental to segmentations. Therefore, we tested aiSEGcell’s robustness against focus drifts. We imaged murine granulocyte/monocyte progenitors (mGMP) with focal planes in the range of ±10 μm (step-size 0.2 μm) equally distributed around the optimal focal plane selected by the experimenter (n = 24 images, N = 1 biological replicate). Segmentations of D1 trained aiSEGcell of all 101 focal planes were compared to the ground truth segmentation in the optimal focal plane. The was 0.88 versus >0.80 for the optimal focal plane versus a range of 4.4 μm (-1.8 μm to +2.6 μm). This range approximately corresponds to the radius of most primary hematopoietic cells, suggesting that aiSEGcell can accurately segment nuclei in BF as long as the focal plane aligns with parts of the nucleus and demonstrating its robustness to focus changes (S2 and S6 Figs and S7 and S8 Tables).

The diversity of BF imaging data will inevitably lead to use cases to which aiSEGcell trained on D1 does not generalize well. For example, megakaryocytes have large, lobated nuclei that are morphologically very different from the nuclei observed in D1 or D2. Hence, we tested the feasibility of adapting aiSEGcell pretrained on D1 to a data set containing murine megakaryocytes (mMEG) co-cultured with murine hematopoietic stem cells (mHSC; D3; Figs 3d and S7a and S9 and S10 Tables). To prevent overfitting on a single experiment, we used one separate biological replicate for training, validation, and test set, respectively (Table 1). We then retrained aiSEGcell with a small set of 34 images, realistic for a typical experimental laboratory use case, to improve mMEG nucleus segmentations. Whereas the pretrained model could detect nuclei of mHSCs, it failed to segment the much larger nuclei of mMEGs (nucleus mask > 750 μm²; Figs 3e and S2c). Retraining aiSEGcell significantly improved segmentations of mMEG nuclei from an average image-wise IOU of 0.02 to 0.77 (pre- / retrained s.d. = 0.05 / 0.08; Fig 3f and S11 Table). The average image-wise IOU of small cells (i.e. mHSCs; nucleus mask ≤ 750 μm²) improved from 0.38 to 0.64 due to the training data containing both large and small cells (S8 Fig). Following the same strategy, we compiled a data set of adherent non-round murine embryonic stem cells (mESC; D4; Table 1). While the pretrained model failed to segment mESC nuclei, retraining aiSEGcell with 32 images enabled it to segment mESC nuclei in BF images, significantly improving the from 0.01 to 0.93 (pre- / retrained s.d. = 0.01 / 0.02; Figs 3d, 3g, 3h, S2d, and S7b and S12–14 Tables). Of note, retraining aiSEGcell on D3 or D4 decreased its segmentation accuracy on D2 (D1 trained / D3 retrained / D4 retrained = 0.66 / 0.27 / 0.12; S3 Table). Training aiSEGcell with only 32 images on a mid-class graphics processing unit (i.e. NVIDIA TITAN RTX) for 1,500 epochs (i.e. iterations over the 32 images) took approximately four hours.

Segmentations with aiSEGcell are based on BF images rich in information about cells, cell interactions, and sub-cellular structures. Understanding which information in BF images is crucial for accurate nucleus segmentations can help efficiently assembling data sets for retraining. Therefore, we selected 999 cells from D2 that aiSEGcell segmented accurately (i.e. high IOU) and 951 cells from D2 that aiSEGcell segmented poorly (i.e. low IOU). Next, we extracted 21 conventional shape and intensity features that describe the nucleus, the whole cell, and the cell neighborhood given the corresponding segmentation masks (S9 and S10 Figs and S15 Table). Taken individually, extracted features were not indicative of whether a cell was segmented accurately or not (S11 Fig). However, training a random forest to classify if a cell is segmented accurately or not (test accuracy = 0.85) revealed that features associated with the nucleus of a cell were more important than features of the whole cell or the cell neighborhood (S12 Fig) [27]. We visually confirmed the importance of nuclear features in BF images by directly interpreting aiSEGcell segmentations with Grad-CAM [28] (S13 Fig).

Removing a fluorescence channel dedicated to image segmentation enables one or even several additional fluorescent markers and thus improved experimental workflows. For example, reporters like GFP are common in transgenic mouse models and are not optimized for multiplexing (i.e. overlapping with cyan and yellow channel). Thus, removing the need for imaging GFP enables the use of both cyan and yellow channels. To illustrate the possibility of analyzing additional biological information in the fluorescent channel freed up by aiSEGcell, we replaced the previously used H2BmCHERRY nuclear marker [25] by virally transducing with fluorescently tagged proliferating cell nuclear antigen (PCNA) to enable live cell cycle phase detection. We isolated lineage unbiased mGMPs (mGMP^GM) from an existing GFPp65 reporter mouse line and utilized D1-trained aiSEGcell nucleus segmentations to jointly image PCNA, p65, lymphocyte antigen 6 family member C1 (Ly6C), and CD115 (D6; Fig 4a) [29]. We tracked cells for two generations to quantify NfκB signaling dynamics following TNFα stimulation (p65), cell cycle phase states (PCNA), and the differentiation into M versus G lineage committed mGMPs (CD115 versus Ly6C expression; Fig 4b) [30,31]. First, we quantified the distribution of NfκB response types in mother cells. In the control condition, 96% versus 4% of mGMP^GM were NON versus TRA (Kull et al.: 96% NON, 0.4% SUS, 4% TRA). In contrast, 81% (15% OSC, 65% TRA) versus 19% of mGMP^GM were responsive to TNFα, confirming results from Kull et al. (9% NON, 13% OSC, 2% SUS, 76% TRA; Fig 4c) [25]. Next, we tested if mGMP^GM respond differently when stimulated in G1- or S-phase. Irrespective of the cell cycle, TRA responders (49% G1, 57% S) were more frequent than NON (37% G1, 38% S) and OSC responders (14% G1, 5% S; Fig 4d). This trend held true when considering only TNFα stimulated mGMP^GM, yielding no credible evidence of the cell cycle affecting NfκB responses to TNFα (G1: 18% NON, 18% OSC, 63% TRA, S: 24% NON, 6% OSC, 71% TRA; S14 Fig). Murine GMPs divide into three sub-populations (mGMP^GM (Ly6C-CD115-), granulocyte lineage biased mGMPs (Ly6C+CD115-), monocyte lineage biased mGMPs (mGMP^M; Ly6C+CD115+)) with characteristic NfκB responses that affect cell differentiation [25,31]. Therefore, we quantified the frequency of cell trees that turned on lineage markers Ly6C versus CD115 depending on the NfκB response type. NON cell trees predominantly remained lineage unbiased, whereas OSC and TRA signaling dynamics significantly enriched for Ly6C-CD115+ cell trees (NON / OSC / TRA 87% / 33% / 42% Ly6C-CD115-, 3% / 0% / 0% Ly6C+CD115-, 8% / 58% / 54% Ly6C-CD115+, 3% / 8% / 4% Ly6C+CD115+; Fig 4e). This pattern remained consistent when correcting for the cell cycle, suggesting that mGMP^GM primarily differentiate into mGMP^M upon TNFα stimulation (Fig 4f). Importantly, these experiments show that aiSEGcell can increase marker multiplexing.

Fig 4. aiSEGcell permits quantification of an additional fluorescent marker.

Nuclear marker H2BmCHERRY was replaced by aiSEGcell to free the fluorescent channel for cell cycle marker mRUBY2PCNA. (a) BF, p65, PCNA, Ly6C, and CD115 images of a representative cell tree. The cell at 23.6 h is a daughter of the cell at 1 h and 4.5 h (time points as grey vertical lines in (b)). Contrast was individually adjusted here to improve visibility (scale bars 10 μm). (b) Temporal information available for a representative cell tree (n = 101 cell trees, N = 2 biological replicates). Cell cycle phases were derived from PCNA. Mother cell NfκB signaling dynamics were classified into four response types: NON, OSC, SUS, TRA. Cell trees were classified as either positive (+) or negative (-) for Ly6C and CD115, respectively (dashed black lines are threshold). The magenta box illustrates a single TNFα addition after 1 h. Trace of upper daughter cell shown in blue. (c) NfκB response type frequencies were comparable to literature reference in the control/stimulation conditions (-TNF / +TNF; ours n = 23 / 78 cell trees, Kull et al. n = 268 / 203, one-sided Chi² test, degrees of freedom = 22 / 77, Chi² statistic = 0.1 / 11.2, p = 0.947 / 0.011; ***: p<0.001, *: p<0.05, n.s.: p≥0.05) [25]. (d) No credible evidence of cell cycle (CC) phase at stimulation affecting NfκB response type frequencies in mGMP^GM. Signaling dynamics of control and stimulation condition are shown. See S14 Fig for only stimulated cells. One outlier was removed due to weak PCNA signal (one-sided Chi² test, degrees of freedom = 20, Chi² statistic = 1.5, p = 0.464). (e) OSC and TRA NfκB responders enriched for Ly6C-CD115+ (-/+) progeny (Ly6C-CD115-: -/-; Ly6C+CD115-: +/-; Ly6C+CD115+: +/+; NON / OSC / TRA n = 37 / 12 / 50; one sided Chi² test, NON vs. OSC / NON vs. TRA / OSC vs. TRA: degrees of freedom = 11 / 49 / 11, Chi² statistic = 43.0 / 143.0 / 0.8, p = 2.5e-9 / 8.6e-31 / 0.664). Two outlier cell trees were removed due to confounding Ly6C or CD115 signal. (f) Enrichment of Ly6C-CD115+ cells in OSC and TRA NfκB responders was not mediated by cell cycle phase at stimulation. Three outlier cell trees of (d) and (e) were removed (n = 98 cell trees).

https://doi.org/10.1371/journal.pcbi.1012361.g004

aiSEGcell is an open source software accessible through a CLI (https://github.com/CSDGroup/aisegcell) or a GUI (napari plugin, https://github.com/CSDGroup/napari-aisegcell) [17]. The CLI version can be used to train, test, and predict with aiSEGcell. Training a neural network is challenging and may require tweaking many different parameters for good performance. Therefore, we provide extensive documentation and notebooks to help users getting started with our tool. We asked a group of 5 testers composed of computationally in- and experienced users to install and use the CLI version based only on its online documentation. All testers successfully installed the CLI version within 15 minutes and were able to train, test, and infer with aiSEGcell on macOS, Windows, or Ubuntu. Finally, we provide users with two pretrained models. The model trained on D1 (S1 Fig), and a model trained to segment whole cells in BF.

Previously published transgenic mice were used for D1, D2, D6 and C57BL/6j mice were used for D3 [25,29]. Experiments were conducted with 10-16-week-old (D1, D2, D6) or 8–12-week-old mice (D3) and only after mice were acclimatized for at least 1 week. Mice were housed in improved hygienic conditions in individually ventilated, environmentally enriched cages with 2–5 mice per cage and ad libitum access to standard diet and drinking water. Mice were housed in a humidity (55 ± 10%) and temperature (21 ± 2 °C) controlled room with an inverse 12 h day-night cycle. Animal facility caretakers monitored the general well-being of the mice by daily visual inspections. Mice displaying symptoms of pain and/or distress were euthanized. Mice were randomly selected for experimental conditions and in some cases pooled to minimize biological variability.

GFPp65 and H2BmCHERRY transgene expression was verified by flow cytometry. Mice were homozygous for GFPp65 (D6; Fig 4) or homozygous for GFPp65 and heterozygous for H2BmCHERRY transgenes (D1-D2; Figs 1 and 2 and S16 Table) [25,29].

Primary cells were isolated as previously described [25,41]. In brief, femora, tibiae, coxae, and vertebrae were extracted, ground in phosphate buffered saline (PBS; 2% fetal calf serum (FCS), 2 mM ethylenediaminetetraacetic acid (EDTA)) on ice and filtered through a 100-μm nylon mesh. Cells were depleted for erythrocytes for 3 min on ice using ammonium-chloride-potassium (ACK) lysing buffer (Gibco) and stained with biotinylated anti-CD3ε, -CD19, -NK1.1, and -Ly-6G (murine monocytic progenitors) or biotinylated anti-B220, -CD3ε, -CD11b, -CD19, -Ly-6G, and -TER-119. For D6 experiments, biotinylated anti-CD41 was also included in the staining. Streptavidin-conjugated magnetic beads (Roti-MagBeads, Roche) were added for immuno-magnetic separation. GMP subpopulations (CD16/32-PerCy-Cy5.5, CD115-BV421, Sca1-BV510, streptavidin-BV570, cKit-BV711, CD34-eFL660 and Ly6C-APCfire), multipotent progenitor populations (CD135-PerCp-eFl710, Sca1-PacBlue, cKit-BV510, CD150-BV650, streptavidin-BV711, CD34-eFluor660 and CD48-APCeFl780), pre-granulocyte/monocyte progenitors (mPreGM), megakaryocyte progenitors, pre-megakaryocytic/erythroid progenitors (CD41-PerCp-eFl710, CD105-eFl450, Sca1-BV510, streptavidin-BV570, CD150-BV650, cKit-BV711 and CD16/32-APC-Cy7), monocytic progenitors (streptavidin-BV650, CD115-BV-421, cKit-BV711, CD135-APC, CD11b-BV605, Ly6C-BV510), mGMP, and mHSC (CD16/32-PerCp-Cy5.5, Sca1-PacBlue, cKit-BV510, CD150-BV650, streptavidin-BV711, CD34-eFluor660 and CD48-APCeFl780) were stained for 90 min on ice, and sorted in single cell mode and sorting purities ≥ 98% with a 70-μm nozzle, using a BD FACS Aria I or III. For D6 experiments, mPreGMs were stained (streptavidin-BV570, CD16/32-PerCy-Cy5.5, CD115-BV421, cKit-BV510, CD150-BV650, Sca1-BV711, Ly6C-APC, CD105-APC/Fire 750) for 90 min on ice and isolated with a 100-μm nozzle in single cell mode and sorting purities ≥ 98% using a BD FACS Aria III. Murine PreGMs were virally transduced for 36 h, stained with the same panel on ice for 90 min to sort mGMP^GM positive for monomeric RUBY2 (mRUBY2) with a 100-μm nozzle in single cell mode and sorting purities ≥ 98% using a BD FACS Aria III (S15 Fig and S17 Table).

Cells were isolated as previously described [42–44]. Femora, tibiae and vertebrae were dissected and crushed in PBS (2% FCS, 2 mM EDTA) on ice. After filtration through a 100-μm nylon mesh, erythrocytes lysis with ACK lysing buffer, and lineage depletion with EASYSep mouse hematopoietic progenitor cell isolation kit (STEMCELL Technologies), cells were stained with anti-CD34-eFluor450, -Sca1-PerCP-Cy5.5, -CD48-FITC, -cKit-PE-Cy7, -CD135-PE-CF594, -CD150-PE, -CD41-APC, and streptavidin-APC-eFl780 for 90 min on ice. mMEGs and mHSCs were subsequently sorted with a BD FACS Aria III (100-μm nozzle, 4-way purity; S17 Table).

Lentiviruses were produced and transduced as previously described [45]. In brief, PCNA was fused to fluorescent protein mRUBY2 and cloned into vesicular stomatitis virus glycoprotein pseudotyped lentivirus (third generation) constructs using the In-Fusion Cloning system (Takara Bio) [46–48]. The virus was produced in human embryonic kidney 293T cells and titrated using NIH-3T3 fibroblasts. Lentivirus was added at a multiplicity of infection of 50–100 to fluorescence-activated cell sorting (FACS) purified mPreGMs in a round bottom 96-well plate for 36 h (37°C, 5% O2, and 5% CO2) before purifying transduced cells by FACS.

Primary cells of experiments 1–7, 13–15, 18–19, and 27–28 were cultured as previously described [25]. After isolation as described above, cells were seeded in 4-well micro-inserts (ibidi) within a 24-well plate with glass bottom (Greiner) and settled for 30–90 min. Before seeding, the plate was coated with anti-CD43-biotin antibody (10 μg/mL in PBS) for 2 h at room temperature and washed with PBS [49]. After cells had settled, 1 mL of IB20/SI media (IB20 = custom IMDM without riboflavin (Thermo Fisher) supplemented with 20% BIT (STEMCELL Technologies), 50 U/mL penicillin (Gibco), 50 μg/mL streptomycin (Gibco), GlutaMAX (Gibco), 2-mercaptoethanol (50 μM; Gibco); SI = 100 ng/mL murine SCF (Peprotech), 10 ng/mL murine IL-3 (Peprotech)) was added per well and cells were cultured at 37°C, 5% O2, and 5% CO2. For D6, Ly6C-APC (1:1,000) and CD115-BV421 (1:1,000) were added to IB20/SI media (S17 Table) [50,51].

Common murine macrophages of experiments 8–12 and 16–17 were obtained as previously described [41]. Common monocytic progenitors (Lin^negCD115^posCD117^posCD135^neg, see above) were cultured in MEM alpha (Gibco) with 10% fetal bovine serum (PAA), 1% penicillin/streptomycin, 50 ng/mL M-CSF (Peprotech), 1% GlutaMAX, at pH 6.9. Following a 3-day incubation at 37°C, 5% O2, and 5% CO2, the cells were stimulated with culture media containing 100 μg/mL ascorbic acid (Sigma-Aldrich), with or without 100 ng/mL Receptor activator of NFκB ligand (RANKL; Miltenyi), to induce osteoclast or monocyte-macrophage lineage, respectively. Differentiation into macrophages was morphologically verified (S17 Table).

Murine ESCs were cultured as previously described [52]. In brief, mESC lines were passaged every 2 days in serum + leukemia inhibitory factor (LIF) containing medium (DMEM (Thermo Fisher), 2 mM GlutaMAX, 1% NEAA (Thermo Fisher), 1 mM sodium pyruvate (Sigma-Aldrich), 50 μM 2-mercaptoethanol, 10% FCS, 10 ng/mL LIF (Cell Guidance Systems)) on gelatin (0.1%; Sigma-Aldrich). Before experiments, cells were cultured for minimum 6 days in 2iLIF medium (NDiff227 (Takara Bio), 1 μM PD0325901 (Selleckchem), 3 μM CHIR99021 (R&D Systems)) on e-cadherin-coated (STEMCELL Technologies) cell-culture plates. All cell lines were tested negatively for mycoplasma using PCR.

Experiments 1–7 and 15 have been published and were acquired as previously described [25]. Experiments 8, 9, 12, 16, and 18 were acquired using Nikon NIS acquisition software on a Nikon A1 confocal microscope with a Nikon A1-DUG-2 detector with a 20x/0.75 CFI Plan Apochromat λ objective. Experiments 10, 11, 13, 14, 17, and 19 were acquired using Nikon NIS acquisition software on a Nikon W1 SoRa Spinning Disk confocal microscope with a Hamamatsu Flash 4.0 or a Hamamatsu ORCA-Fusion camera, 20x/0.75 CFI Plan Apochromat λ objective, and Spectra X (Lumencor) light source for BF. Media conditions for murine monocytic progenitors/macrophages and other primary hematopoietic cells were as described above. Images were acquired every 6–9 min and movie lengths varied between 12–48 h. After 1 h, the movie was briefly stopped to stimulate cells with control media, TNFα- (Peprotech), lipopolysaccharide- (Sigma-Aldrich), or RANKL- (Miltenyi) supplemented media. All time lapse experiments were conducted at 37°C, 5% O2, and 5% CO2 (S16 Table).

Time lapse experiments were conducted as previously described [22,48]. In brief, mHSCs and mMEGs were co-cultured at 37 °C, >98% humidity, 5% O2, and 5% CO2 on a μ-slide (ibidi), coated with 5 μg/mL anti-CD43, in phenol-red-free IMDM media (20% BIT, 100 ng/mL SCF, 100 ng/mL thrombopoietin (R&D Systems), 10 ng/mL IL-3, 1x GlutaMAX, 50 μM 2-mercaptoethanol, 50 U/mL penicillin, 50 μg/mL streptomycin) for 24 h before imaging on a Nikon Eclipse Ti-E equipped with linear-encoded motorized stage, 20x/0.75 CFI Plan Apochromat λ objective, Hamamatsu Orca Flash 4.0 V2, and Spectra X fluorescent light source (Lumencor; S16 Table).

First, 3,000 mESCs were seeded per channel of an E-Cadherin coated μ-slide (ibidi) in NDiff227 medium containing doxycycline (Dox; 1 ng/mL; Sigma-Aldrich) and overlaid with silicone oil. After approximately 20 h, cells were washed 5x to remove Dox and the medium was replaced with NDiff227 containing different combinations of the small molecules Dox, trimethoprim (Sigma-Aldrich), and 4-hydroxytamoxifen (Sigma-Aldrich). Imaging was performed on Nikon Eclipse Ti-E microscopes equipped with 10x/0.45 CFI Plan Apochromat λ objective, Hamamatsu Orca Flash 4.0 cameras, and Spectra X fluorescent light source (Lumencor). Images were acquired using custom software [53]. All time lapse experiments were conducted at 37°C, 5% O2, and 5% CO2. Only a single time point of the time lapse experiment was used for all experiments involving D4 (S16 Table).

Experiment 26 on hMPPs (S16 Table) has previously been published [22]. Signaling dynamics derived from aiSEGcell segmentations were processed and quantified with the respective analysis pipeline.

Experiments were acquired using Nikon NIS acquisition software on a Nikon W1 SoRa Spinning Disk confocal microscope with a Hamamatsu ORCA-Fusion camera, 20x/0.75 CFI Plan Apochromat λ objective, and Spectra X (Lumencor) light source for BF. Media conditions for mGMP^GM were as described above. Images were acquired every 9 min for 36 h. After 1 h, the movie was briefly stopped to stimulate cells with control media or TNFα-supplemented media (40 ng/mL final concentration). All time lapse experiments were conducted at 37°C, 5% O2, and 5% CO2 (S16 Table).

Images were processed and quantified as previously described [5,25,48,54–59]. In brief, 12-bit or 16-bit images were saved as TIFF-files and linearly transformed to 8-bit PNG-files using channel-specific black- and white-points. Fluorescence channels were background corrected for D3 and D4. Images were segmented and individual cells of D2 and D6 were tracked for subsequent fluorescence channel quantification. Missing values due to missing masks were mean imputed. For D2 NfκB normalization, mean nuclear NfκB intensities were computationally resampled in intervals of 7 min using linear interpolation to account for the different experimental imaging frequencies. Mean NfκB intensity was divided by the cell-wise average of mean NfκB intensity prior to stimulation and rescaled to a range of 0 and 1 using experiment-wise min-max scaling. For D6 NfκB normalization, mean nuclear NfκB intensity was divided by mean whole cell NfκB intensity. Normalized NfκB signal was divided by the cell-wise average of normalized NfκB signal prior to stimulation to obtain baseline normalized NfκB signal used for subsequent analysis. Cell cycle phases were manually assigned during cell tracking. Ly6C and CD115 status was assigned by thresholding on the whole cell mean intensity divided by the cell-wise average of whole cell mean intensity prior to stimulation. For Ly6C, cell-wise mean and standard deviation prior to stimulation were computed to set threshold = mean+5 * s.d.. For CD115, cell-wise mean prior to stimulation were computed to set threshold = 3 * mean. Assignments of Ly6C and CD115 status were manually confirmed. NfκB signaling dynamics were normalized and classified in R (4.0.2).

Ground truth nuclear masks were obtained by manually curating fastER segmentations [5]. In brief, fastER uses human-in-the-loop region annotations to train a support vector machine that segments candidate regions based on shape and texture features and uses a divide and conquer approach to derive an optimal set of non-overlapping candidate regions. fastER was trained experiment-wise on the respective nuclear marker channel and resulting masks were eroded (settings: dilation -2). fastER accurately segmented most nuclei, but light path differences between fluorescence and BF (e.g. shadow obstructing cell in BF) or, for example, focus shifts of live non-adherent cells led to missing or inaccurate object segmentations. Consequently, inaccuracies in segmentations were manually curated with a custom napari plugin if necessary [17]. aiSEGcell segmentations were obtained by predicting and thresholding (0.5) on a single foreground channel from BF images with the respective trained model and subsequently removing small objects (skimage.morphology.remove_small_objects) and small holes (skimage.morphology.remove_small_holes) with experiment-specific settings. Semantic segmentations were converted to instance segmentations using connected components (skimage.measure.label) [60]. Whole cell segmentation masks for D6 experiments were obtained using the aiSEGcell model to segment whole cells in BF accessible through CLI and GUI. For D5 and D6, aiSEGcell segmentations were manually curated in BF if necessary [55].

Murine GMPs were isolated from C57BL/6j mice as described above. 10,000 mGMPs were seeded per well of a micro-Insert 4 Well (ibidi) and were stained with SPY555-DNA (1:1,000; Spirochrome) for 300 min at 37°C, 5% O2, and 5% CO2. The experiment was acquired using custom software [53] on a Nikon Eclipse Ti2 with a Hamamatsu ORCA-Fusion camera with a 20x/0.75 CFI Plan Apochromat λ objective, and Spectra X (Lumencor) light source (S16 Table).

Murine GMPs were isolated from a C57BL/6j mouse as described above. Cells were stained with SPY650-DNA (1:1,000; Spirochrome) for 60 min on ice. The experiment was acquired using Nikon NIS acquisition software on a Nikon W1 SoRa Spinning Disk confocal microscope with a Hamamatsu ORCA-Fusion camera with a 20x/0.75 CFI Plan Apochromat λ objective, and Spectra X (Lumencor) light source for BF. The experimenter manually selected the optimal focal plane and focal planes in the range of ±10 μm (step-size 0.2 μm) equally distributed around the optimal focal plane were acquired for 24 field-of-views. Media conditions for mGMPs were as described above. Images were processed as for D1. Ground truth segmentations were obtained using fastER on the nuclear fluorescence signal of the optimal focal plane, (D1-trained) aiSEGcell segmentations were obtained for all focal planes [5].

NfκB signaling dynamics were classified as NON, OSC, SUS, TRA, or unclear/outlier by the experimenter. The experimenter was blind to the experimental conditions of the randomly displayed signaling dynamics and responder-, OSC-, and SUS-TRA-scores were displayed [25].

We quantified segmentation performance with two types of F1-scores. The conventional F1-score was implemented as previously described [61]. In brief, object-wise IOUs were computed for each object in a ground truth segmentation mask and the predicted segmentation mask. Ground truth objects with an IOU>τ₁ were considered a true positive (TP). Ground truth objects with no matching predicted object (IOU≤τ₁) and predicted objects with no matching ground truth object were considered FN and FP, respectively. We then computed , with ε a small numerical term to prevent division by 0, for varying thresholds τ₁ (0.5–0.9 in 0.05 increments), Conventional F1-scores were provided in S2, S3, S5, S6, S8, S10, S11, S13, and S14 Tables.

We observed that most errors were nuclei segmented with only small inaccuracies, which were nevertheless counted as FP and FN. This led the conventional F1-score to underestimate the segmentation quality (S16 Fig). Therefore, we introduced a second threshold (τ₂ = 0.1) to prevent counting inaccurately predicted objects (e.g. IOU = 0.4) as both, FP, and as FN. FPs were objects in the prediction with IOU≤τ₂ and FNs were objects in the ground truth with IOU≤τ₂. We introduced the category of inaccurate masks (IA), ground truth objects with exactly one τ₂<IOU≤τ₁ and computed the adapted image-wise . Given the small size of primary cell nuclei at 20x magnification (many nuclei <100 pixels) and the sensitivity of the F1 computation to object size, we observed τ₁ = 0.6 to be an appropriate threshold and denoted the image-wise average adapted F1-score with τ₁ = 0.6 as throughout this manuscript [21,61,62]. Analogous to the conventional F1-score, we provided the adapted F1-score for varying τ₁ thresholds (S1, S3, S5–S7, S9, S11, S12, and S14 Tables). Ground truth objects with τ₂<IOU≤τ₁ for more than one predicted object were considered a split. Predicted objects with τ₂<IOU≤ τ₁ for more than one ground truth object were considered a merge.

U-Net was selected as model architecture due to its wide and robust applicability in biological image segmentation [16,63]. Leaky ReLU and batch normalization were used in convolutional blocks [64,65]. Two convolutional blocks and a maximum pooling layer or bilinear upsampling layer formed the down- and upscaling blocks, respectively. Convolutional kernels (3×3) were padded, and the number of kernels doubled with each downscaling block (capped at 512), starting at 32. The number of kernels halved with each upscaling block. In total, 7 downscaling blocks were used to result in a receptive field of 128×128 pixels for the lowest layer kernels. Model instances were trained on randomly augmented (flipping, rotation) 1,024×1,024-pixel images or random crops of 1,024×1,024 pixels from larger images. For D3 and D4 retraining, random blurring/sharpness augmentations were additionally used to improve training on the very small and less diverse training data. Binary cross entropy was used as loss function.

For D1 training, model instances were trained for 400 epochs with a batch size of 14, the maximum batch size to fit on two NVIDIA TITAN RTX GPUs during training. The weight of the loss function was 1. Model instances were initialized with random weights and three replicates for each learning rate (5e-3, 1e-3, 5e-4, 1e-4) were trained [66]. The best model was selected based on the highest average image-wise F1-score on the validation set. For D3 retraining, model instances were initialized with the weights of the best model instance trained on D1 and were trained for 1,500 epochs with a batch size of 4 to balance generalizability and the small training set size. The learning rate was fixed at 1e-5 and three replicates for each loss weight (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 50, 100, 150, 200) were trained. The best model was selected based on the highest average image-wise IOU for big cells on the validation set. For D4 retraining, model instances were initialized with the weights of the best model instance trained on D1 and were trained for 1,500 epochs with a batch size of 4. A grid search to find the best model instance considered learning rate (5e-5, 1e-5, 5e-6, 1e-6) and loss weight (1, 2, 3) with three replicates for each hyperparameter combination. The best model was selected based on the highest average image-wise F1-score on the validation set.

We used Cellpose version 3.0.7 and selected “nuclei” as a generalist nucleus segmentation model and “yeast_BF_cp3” as a bright field round object segmentation model [20]. A grid search to find the best post-processing hyperparameters considered diameter (17, 30, 40) and flow threshold (0.2, 0.4, 0.6).

We used StarDist version 0.8.5 and selected available pretrained models “2D_versatile_fluo” and “2D_versatile_he” for testing [19]. A grid search to find the best post-processing hyperparameters considered probability threshold (0.5, 0.8, model-specific default) and non-maximum suppression threshold (0.2, 0.4, model-specific default).

We selected 3,053 images of D2 that did not cause an error during feature computation (i.e. cell contour identification) due to too close cell masks. In this subset of D2, for all experiments selected the 200 cells with the highest IOU and 200 cells with the lowest IOU. After manual curation 1,950 cells (999 high, 951 low) remained, for which nucleus and whole cell masks were available (aiSEGcell whole cell segmentation). The 21 shape and intensity features were computed as described (S15 Table) [60,67].

The best random forest classifier was selected by a grid search over training set size (50, 100, 200, 300, 400, 500, 600, 700, 800, 1,000), number of estimators (20, 50, 100, 200, 300, 400, 500, 600), and maximum estimator depth (5, 7, 9, 11, 13) with 4-fold cross validation. Training and test sets were balanced, and cells were randomly selected from the 1,950 available cells. The best model had 600 estimators with a maximum estimator depth of 7 and was trained on 200 cells. Feature importances were based on the Gini importance. The scikit-learn implementation of random forest was used [68].

GradCAM heatmap overlays were obtained using the python grad-cam package [69].

All data visualization and statistical analyses were conducted in Python (3.10.9). When appropriate, distributions were tested for normality (Shapiro-Wilk test) to select the statistical test for group comparisons [70].