DeepIFC: Virtual fluorescent labeling of blood cells in imaging flow cytometry data with deep learning

Imaging flow cytometry (IFC) combines flow cytometry with microscopy, allowing rapid characterization of cellular and molecular properties via high‐throughput single‐cell fluorescent imaging. However, fluorescent labeling is costly and time‐consuming. We present a computational method called DeepIFC based on the Inception U‐Net neural network architecture, able to generate fluorescent marker images and learn morphological features from IFC brightfield and darkfield images. Furthermore, the DeepIFC workflow identifies cell types from the generated fluorescent images and visualizes the single‐cell features generated in a 2D space. We demonstrate that rarer cell types are predicted well when a balanced data set is used to train the model, and the model is able to recognize red blood cells not seen during model training as a distinct entity. In summary, DeepIFC allows accurate cell reconstruction, typing and recognition of unseen cell types from brightfield and darkfield images via virtual fluorescent labeling.


| INTRODUCTION
Imaging flow cytometry (IFC) is a recent technique which combines fluorescent microscopy and flow cytometry into a high-throughput analysis platform [1].IFC allows for the study of cellular and molecular properties in fluidic samples at a single-cell level in high-throughput manner.It has been found useful in quantifying nucleic acids and protein expression [2], classifying rare cell types [2], identifying cells in their early apoptotic stage [1], examining host-intracellular parasites [3] and other diagnostic purposes in hematology [4].In recent years, machine learning on IFC data [5] has been used to predict DNA content, quantify mitotic cell cycle phases [6], reconstruct diabetic retinopathy disease progression and the cell cycle of Jurkat cells [7] as well as to classify and identify white blood cells [8,9].
Fluorescent labeling requires a considerable amount of time, resources and effort, and can damage the cells [10].Consequently, socalled label-free or virtual staining approaches have been considered which may allow bypassing fluorescent labeling altogether.Recent studies in fluorescent imaging have used deep learning models to virtually stain brightfield images of adipose tissue [11], detect acute lymphoblastic leukemia cells [12] and to discriminate between different cell lines [13].Machine learning on IFC brightfield and darkfield images have been used to distinguish cell types [8] and transitions between cell states [7].
Label-free deep learning methods have been created to reconstruct fluorescent images from brightfield images [14][15][16], but to our knowledge, the reconstruction of single-cell multichannel fluorescent images in IFC data has not been proposed.Label-free cytometry methods based on segmentation and unsupervised modeling [15], weak supervision [17], cytometry by time of flight (CyTOF) [18] and time-stretch microscopy [19] have been suggested.Previous methods have also utilized the Amnis IDEAS ® software analyses, such as nuclear localization, or user generated cell masks [8], which allow for the segmenting of the cell from its surroundings based on pixel intensity.Cell masks may negatively impact analysis, if their accuracy is not high enough [20].They may also hinder recognizing the differences between cell states [21].Spatial distribution of labels and dim-bright label continuum are also not explicitly modeled in these approaches.
In this study we present a novel method called DeepIFC (Figure 1
In addition to these surface markers, images exhibiting positivity for 7-AAD were successfully reconstructed (r = 0.79, AUROC = 0.954).To evaluate the ability of DeepIFC to identify cell types in the complete PBMC data, we assigned cell types based on thresholded fluorescent intensities in images generated by DeepIFC (strict cell typing strategy; Methods).We also performed the same cell typing procedure for observed images by manual gating in IDEAS ® software (Supplementary Figure 4a) as well as thresholding the mean intensity of each fluorescent image (Supplementary Figure 4b) to establish two different ground truth settings.When compared to the image-based ground truth, DeepIFC was able to accurately classify monocytes (90% recall, 78% precision) likely due to their distinct morphology (Figure 2B, Supplementary Table 3).Despite both CD3 and CD45 markers being well predicted individually, performance predicting CD8-CD56-T cells was found to be at the moderate level (80% recall, 73% precision).This was mostly due to falsely predicted CD8 and 7-AAD label fluorescence, since 6% and 7% of CD8-CD56-T cells were predicted to be cytotoxic T cells, or dead or damaged cells, respectively.Dead or damaged cells were predicted at 73% recall and 54% precision.DeepIFC showed high performance of 87% recall and 92% precision when predicting cells of unknown type, that is, cells where the predicted marker fluorescences did not correspond to any known combination or are all negative (Figure 2B, Supplementary Figure 4b,c).In contrast to these well-predicted types, the cell types appearing in smaller amounts in the data or exhibiting T cell subtype markers were predicted at much lower recall levels (NK, 25%; NKT, 24%; cytotoxic T, 13%; B, 5%).Unsurprisingly, the most common incorrect prediction for NKT cells was a T cell, with 41% of true NKT cells classified as T cells.Likewise, 72% of true cytotoxic T cells were classified as T cells.Of true NK cells, 31% were identified as WBCs due to failure to predict CD56 fluorescence from morphology.The predictive performance on multiple cell types improved when dead or damaged, unknown cells and cells positive for only the CD45 marker were removed from analysis.Most notably, recall and precision for monocytes improved from 90% and 78% to 98% and 92%, respectively (Supplementary Table 3).Cell type fractions predicted by Dee-pIFC were found to correspond well to fractions obtained from observed images and analysis in IDEAS ® (Figure 3, Supplementary Table 1).
We then extracted features learnt by the DeepIFC model for each cell in the complete PBMC dataset and visualized them by projecting the features to two dimensions with UMAP [23].We observed four distinct large clusters (  4b).Label "WBC" denotes a white blood cell only exhibiting the CD45 marker, and "Unknown" a cell exhibiting an unknown combination of markers not matching any cell type on the panel, or negativity for all markers.markers in the data did not constitute separate clusters, we found NKT cells to be concentrated to the bottom left of cluster 1, while NK and B cells were found predominantly in the bottom right of cluster 1.Similar clustering of cytotoxic T cells was not found.NK cells were mostly located at the outer edge of the T cell cluster 1 and the dead cell cluster 3, but some were also found in the monocyte cluster 2 (Supplementary Figure 10).

| DeepIFC recognizes a cell type not seen during training
To understand whether DeepIFC models would be useful in analyzing cell types not seen during training, we processed brightfield and darkfield images of red blood cells from a recent IFC study (Reference [24]; dataset Mixed_CE47_D2) with the DeepIFC model trained on PBMC images ("complete dataset") without retraining the model on RBC images, and computed the DeepIFC features for the RBC images.Surprisingly, a large fraction of these RBCs (97%) formed a new cluster (Cluster 6, n = 11,159 cells; Figure 4B) separate from the mononuclear cell clusters, demonstrating the ability of DeepIFC to recognize cells with unseen morphology as a distinct entity.

| Balancing training data improves DeepIFC prediction performance
We then investigated whether it would be possible to improve DeepIFC performance on cell types which were poorly predicted in the complete PBMC dataset.To do this, we created balanced datasets separately for each cell type such that the proportion of the target cell type was set to 50% (Methods).DeepIFC models were trained on multiple balanced datasets with different amounts of target cells to study the effect of number of cells on performance.Prediction performance of DeepIFC models showed substantial improvements on multiple cell types over the performance of models trained with the unbalanced, complete dataset (Figure 5, Supplementary Figure 5, Supplementary Table 3).For most cell types, the saturation point in performance increase was reached at 3200 cells, with the rise in accuracy slowing down with larger cell amounts.Notably, a large performance gain was observed with NKT cells, which were predicted at 83% recall and 63% precision in balanced data compared to 24% recall and 21% precision in unbalanced data.Similarly, recall of NK, cytotoxic T and B cells increased from 25% to 67%, 13% to 68% and 5% to 54%, respectively.Results for average accuracies over three training runs for  5 and Supplementary Table 3.The counts and percentages for the gating of all cell types can be found in Supplementary Tables 1 and 5.

| Performance of DeepIFC on fresh peripheral blood mononuclear cell samples
We sought to understand whether freezing and thawing cells prior to imaging compared to using fresh cells would influence DeepIFC performance.We first evaluated the DeepIFC model trained on the complete dataset obtained from frozen cells by predicting label images in a dataset obtained from four fresh PBMC samples.When compared to frozen cells, fresh cell label predictions ranged from slightly worse (CD45 r = 0.85 fresh vs. r = 0.89 frozen; AUROC 0.97 fresh vs. 0.98 frozen) to substantially worse (CD56 r = 0.17 fresh vs. r = 0.50 frozen; AUROC 0.58 fresh vs. 0.81 frozen) (Supplementary Figure 11, Supplementary Table 7).
Features produced by the model were visualized in the same manner as with the complete model and 2D UMAP representations were created.Distinct clusters for fresh and frozen samples were observed, as well as clusters for cell types, by both models (Supplementary Figure 12a,c,e).Red blood cells were separated in their own cluster, as with frozen sample data.[24].DeepIFC distinguishes RBCs as a distinct entity (cluster 6) based on their morphology despite not seeing large numbers of these types of cells during model training.As red blood cells do not express any marker, they are correctly identified as "unknown" class along with other cells negative for all markers, or with unknown combinations of positive markers.The RBC cluster also connects to the unknown/all negative cluster, but forms a distinct entity.model separated fresh and frozen samples (Supplementary Figure 12b,d).Similar cell type clusters emerged as previously (Supplementary Figure 12f), although in fresh samples, two clusters containing NK cells were formed (Supplementary Figures 12 and 14).
However, when training a new model with both fresh and frozen sample data, predictions did not significantly improve (Supplementary Table 7) and for some labels, such as monocytes, performance deteriorated.

| Prediction of doublet events
The PBMC IFC frozen sample data contained a number of acquisition events where two (doublets) or more cells were imaged at the same time.Scrutiny of fluorescent images generated by the DeepIFC complete data model revealed that the method was able to predict label fluorescence separately for multiple cells in the same event (Supplementary Figure 3).We found the doublet events to contain proportionally fewer T and cytotoxic T cells compared to all events (T, 58.9% of cells in doublet events, 95% CI 54.9-62.7% vs. 65% of cells in all events; cytotoxic T, 10.5%, 95% CI 8.2-13.1% vs. 15%), while exhibiting a two-fold increase in monocytes (24.8%, 95% CI 21.5-28.3%vs. 12%).T cell -T cell doublets were the most common at 23% of all doublet events.Other cell doublets were rarer, for example, cytotoxic T cell -T cell pairing (13% of all doublet events), NKT -T cell (6%), monocyte -T cell (6%) and natural killer -T cell (6%).Monocytes were most often paired with other monocytes (18% out of all doublet events).In fresh sample data, a cluster (n = 1430) of mostly doublet events was observed, including B cell doublets (Supplementary Figure 12f, indicated by an orange arrow).

| DISCUSSION
In this study, we presented a novel computational method called Dee-pIFC for reconstructing fluorescent images from brightfield and darkfield images acquired with an imaging flow cytometer.Although generating images across multiple microscopy modalities has been demonstrated in fluorescence microscopy [25], to our knowledge DeepIFC is the first method employing such cross-modality learning in multichannel IFC.DeepIFC trained on IFC data from PBMCs showed high accuracy in reconstructing fluorescent images for the cell surface markers CD45, CD3 and CD14 solely from morphology.The method was additionally able to accurately recognize dead and damaged cells, which may enable sample quality control in IFC without explicitly staining for damaged cells thus allowing a fluorescent channel to be used for other purposes [1].Interestingly, no distinct difference in dead or damaged cell morphology from live ones was visible to the expert eye in brightfield and darkfield images, despite DeepIFC predicting the fluorescence of dead/damaged cell marker 7-AAD reasonably well.This may be due to the phase of apoptosis where no changes to cell morphology clearly visible to human experts have occurred yet, but the cells already bind 7-AAD.Accurately predicting quality of cells is of critical importance in operation of blood services and biobanks, and thus models such as DeepIFC hold promise to decrease costs and improve throughput in these facilities.
DeepIFC enables identification of cell types and characteristics via virtual gating of the generated fluorescent images.DeepIFC models achieved classification performances ranging from 54% (B cells) to 92% (monocytes) for recall, and from 61% (cytotoxic T cells) to 91% (monocytes) for precision.The reason for the accurate prediction of certain cell types (CD45+ leukocytes, CD3+ T cells, CD14+ monocytes, 7-AAD+ dead/damaged cells) may be that they were either present in high amounts in the data (leukocytes, T cells) or their morphology was distinct from the other cell types (monocytes, dead cells).Cell types appearing in smaller amounts were predicted at much lower recall levels in the complete, unbalanced dataset (NK, 25%; NKT, 24%; cytotoxic T, 13%; B, 5%), however training with balanced data resulted in substantially improved prediction performance compared to unbalanced original data in most cell types (NK, 67%; NKT, 83%; cytotoxic T, 68%; B, 54%) except for CD8-CD56-T cells.
Similarly to previous efforts utilizing machine learning [8], DeepIFC had difficulties distinguishing between T cell subtypes.Morphological differences between T cell subtypes (CD56, CD8) were not found to be visible to the human eye in brightfield images.Regardless, we found DeepIFC to classify the subtypes better than random guess (cytotoxic T, 64% binary classification accuracy; NKT, 70%).
A unique feature of our approach compared to previous IFC data analysis methods which attempt to predict marker positivity as labels [7][8][9] is that DeepIFC reconstructs the entire fluorescent image instead of outputting a binary or class-based prediction.This allows the method to consider the spatial distribution of fluorescence as well as predict fluorescence in multiple cells in the same IFC acquisition event, and offer the user visual cues on the virtual fluorescent labeling process.These capabilities may enable analysis of complex cell-cell interactions without fluorescent labels [26].To this end, DeepIFC was able to identify different cell types in doublet events, highlighting a two-fold increase in the proportion of monocytes in doublet events (25% in doublets vs 12% in all events).Monocyte-monocyte doublets are found in for example, psoriasis patients' blood [27], while T cells are known to form doublets with monocytes in the event of infection [28].
We demonstrated how DeepIFC models can be applied to data not seen during training, for example from cell types not present in training data, to detect novel cell entities.Analyzed with a DeepIFC model trained on data from PBMC samples with RBCs removed, a set of RBCs from an independent study [24] formed a distinct cluster from mononuclear cell types.Machine learning methods able to distinguish labels not present in training datasets (i.e., zero-shot learning) have been studied extensively [25,26].In this study we showed for the first time these capabilities applied to IFC.We envision DeepIFC models trained on larger datasets to be able to distinguish a wide variety of cell types and cellular characteristics.As RBCs have morphology distinct from mononuclear cells, it would be interesting to explore DeepIFC's performance also on other cell types not present in the datasets used in this study.
When tested with IFC data obtained from fresh cells, the Dee-pIFC model trained with frozen samples retained relatively good performance in predicting the labels it had learned well before (CD45, CD3, CD14, 7-AAD), but faced a larger deterioration of accuracy with labels that were already difficult to predict in frozen cells (CD19, CD8, CD56) (Supplementary Table 7).The large number of dead cells in frozen samples may have enabled the model to learn their morphological characteristics, even when not clearly apoptotic.
Interestingly, NK cells were better clustered by DeepIFC when fresh sample data was included in training.It has been found that NK cells may not preserve viability and may lose cytotoxicity after the freeze-thaw process [29].This effect may be reflected in the loss of distinct NK cell morphology, leading to poor predictive performance in frozen samples.In addition, one of the fresh samples included in the study contained an abnormally high amount (31.5%) of NKT cells.The frequency of NKT cells in human blood normally varies from 1.4% to 23% [30], but infections can result in deviations from this normal state [31].
In the future the effect of larger, balanced and augmented datasets may be examined to improve the performance of DeepIFC, as suggested by Lippeveld et al. [8].Cell masks such as those generated by IDEAS ® may reduce the possible bleed-through between channels.
The quality of the samples may also play a role.There was substantial variation in the number of dead cells per sample (6.6%-42.0%) in the frozen and thawed samples.Suboptimal cryopreservation may cause alterations of the cellular phenotype and lead to non-specific binding of antibodies [32,33] as well as changes in morphology.On the other hand, training with fresh samples did not improve the model, potentially due to large differences in morphology and image characteristics between the two datasets.Our approach could be also improved with further optimization of the antibody panel by testing different fluorochromes for different markers and solely training the model with fresh samples.It would also be interesting to study in depth how features associate with the different morphological markers.We leave as future work how to best integrate fresh and frozen cell data for labelfree applications.
Taken together, methods such as DeepIFC able to perform virtual labeling and cell type identification solely from morphology hold promise to transform diagnosis of hematological diseases and blood processing pipelines by not having to introduce fluorescent labels during workflow, reducing costs and processing time required.Possible avenues to develop the method further include utilizing larger training datasets covering more cell types, data augmentation to improve performance on rare cell types [34], and a user-friendly graphical tool to use the software.

| DeepIFC model
We developed a computational workflow called DeepIFC to perform virtual fluorescent labeling on brightfield and darkfield IFC images (Figure 1).For each input cell, DeepIFC workflow results in a generated image for each fluorescent channel, and a set of single-cell features.The generated images are then used to perform virtual gating to classify cell types.DeepIFC also visualizes the single-cell features by projecting the features onto a two-dimensional space with Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) [23].
At the core of DeepIFC is a deep neural network model based on the Inception U-Net architecture [35] (Figure 6).Inception U-Net combines the U-Net architecture common in image segmentation tasks [37] with Inception modules [36] Where Q is the quantile function and q was set to 0.6 in our study, to obtain images similar to those produced by IDEAS ® software (Luminex corporation, Austin, Texas) (Supplementary Figures 1 and 2).

| Isolation and staining of the cells
Human PBMC were extracted from buffy coats of blood samples from six voluntary anonymous blood donors (Finnish Red Cross Blood Service) using Ficoll-Pague™ Plus (GE Healthcare Life Sciences) density gradient centrifugation according to manufacturer's recommendations.PBMCs were frozen to be stained later.We also received four fresh buffy coat samples from random, anonymized donors (MNC1-4, Supplementary Table 5).The PBMC extraction process and fluorescent labels were kept as similar as possible to the original experiment with frozen samples, in order to eliminate any unintended differences in the finished samples.4).Cell samples were also stained with corresponding isotype control antibodies.To reduce the background staining, cells were treated with BD Fc Block (BD Biosciences), following the manufacturer's recommendation.

| Imaging flow cytometry of peripheral blood mononuclear cells
Seven PBMC samples of six blood donors (Supplementary Table 5) were imaged with a 12 channel Amnis ® ImageStream ®X Mark II imaging flow cytometer (ISX) (Luminex/DiaSorin Group) to capture images of mononuclear cells (MNC).One of two samples (NK11B) obtained from the same donor was discarded from the experiment as it contained a high number of cells positive for the damaged/dead cell marker 7-AAD (54%) (Supplementary Table  For the fresh PBMC samples, the fluorescent labels were identical to the previous, frozen experiment, except for darkfield being located in channel 6, and the empty channel 6 being located in channel 12.
We aimed to acquire 100,000 imaging events for each sample.
) to generate fluorescent images solely from morphological information in blood cell IFC data with minimal preprocessing.DeepIFC consists of a deep neural network model trained on brightfield and darkfield images of cells to generate corresponding fluorescent images.In contrast to many other approaches, DeepIFC reconstructs fluorescent images instead of predicting cell class labels such as the cell type, or fluorescent label intensity.The model also learns an intermediate representation of each input cell, which is useful in distinguishing cell types and features.We present tools to examine these representations (i.e., features) visually, and a method to identify cell types based on fluorescent images predicted by the DeepIFC model.Importantly, DeepIFC does not require manually annotated training data (e.g., cell type labels or cell image masks).We trained DeepIFC models on IFC data generated on peripheral blood mononuclear cells (PBMC), and evaluated the performance of these models also on data on red blood cells (RBC), a cell type not used in training.The DeepIFC workflow and models trained on PBMC data are available on GitHub (https:// github.com/timonenv/DeepIFC).

7 -
AAD permeates the cell wall and binds to the DNA sequence of dead or damaged cells.On the other hand, fluorescent images for the surface markers CD56 (r = 0.53, AUROC = 0.805), CD19 (r = 0.61, AUROC = 0.800) and CD8 (r = 0.41, AUROC = 0.728) were less accurately reconstructed, likely due to the relatively small numbers of cells in the complete dataset exhibiting these markers (Supplementary Table 1) or the difficulty in distinguishing these cells only from morphology.As an example of distinct morphology driving label prediction, the monocyte marker CD14 was better predicted than the NK F I G U R E 1 DeepIFC workflow.Imaging flow cytometry images are extracted from compensated image files (CIF) and image background intensities are normalized.Brightfield and darkfield images are processed by the DeepIFC model to generate corresponding fluorescence images.DeepIFC model also yields single-cell features, which the workflow visualizes by projecting the features onto a two-dimensional space with UMAP.The workflow also contains an interactive tool usable in a web browser for displaying the two-dimensional projection together with observed and generated images.[Color figure can be viewed at wileyonlinelibrary.com] and NKT marker CD56 despite both being expressed in similar numbers of cells.In contrast, B cells do not similarly have distinct characteristics from more abundant T cells, resulting in difficulty in predicting the CD19 marker [22].Examples of measured images and images reconstructed by DeepIFC are shown in Supplementary Figures 1 and 2.

Figure 4 )
corresponding to T cells (cluster 1, n = 42,258 cells), monocytes (2, n = 12,292), dead or damaged cells (3, n = 29,548), and objects which did not exhibit any fluorescent label or had unknown combinations of markers (4, n = 12,864).In addition, two smaller clusters consisting of debris were visible (5a, n = 64; 5b, n = 180).A subset of cells (n = 5778; 1%) expressing both the monocyte marker CD14 and dead/damaged marker 7-AAD were found to connect the monocyte and dead/damaged cell clusters.These dead/damaged monocytes are most likely morphologically distinct enough from other cells so that they form a bridge between the two clusters instead of mixing with other dead/damaged cell types (3).Although rarer cell types, that is, cell types exhibiting T cell subtype F I G U R E 2 DeepIFC model predictive performance in PBMC frozen sample data ("complete dataset"), for a test set of 200,000 cells (donors NK10 and NK17).(A) Receiver operating characteristic (ROC) curve for each cell marker.(B) Confusion matrix showing precision for the strict cell typing strategy where the ground truth (target cell types) is provided by observed fluorescent images (Methods, Supplementary Figure

F
I G U R E 3 Cell type fractions obtained from gating strategy performed in IDEAS ® (Supplementary Figure 4a), and thresholding of average fluorescence intensities performed for the ground truth images and DeepIFC predicted images as per the permissible cell typing strategy, for each donor sample.Number of cells indicated for each sample.Comparison of (A) image-based ground truth and IDEAS ® cell type fractions, (B) imagebased ground truth and DeepIFC generated cell type fractions, and (C) DeepIFC generated cell type fractions and IDEAS ® cell type fractions.(D) An inset of the image-based ground truth and DeepIFC comparison restricted to range [0, 40%].each model are shown in Figure 5, and the results for best performing individual models are shown in Supplementary Figure

2. 5 |
Training a DeepIFC model on combined data from fresh and frozen samples A DeepIFC model was trained with both fresh and frozen sample data (Methods).Similarly to the model trained only on frozen cells, the F I G U R E 4 UMAP projections of single-cell features learnt by DeepIFC, for all six frozen donor samples.Features were extracted and combined from the seven DeepIFC models learnt for each cell type in the data.(A) Data from the DeepIFC models trained on complete PBMC data.Cell types identified with the strict typing strategy according to observed images are indicated with colors.Four main clusters are visible: T cells (1), monocytes (2), dead or damaged cells (3), and objects not expressing any fluorescent marker or otherwise unknown (4).Two smaller clusters containing debris are indicated with 5a and 5b.(B) UMAP projection showing combined PBMC and RBC data

F I G U R E 5
DeepIFC binary prediction accuracy (y-axis) in the balanced PBMC datasets with respect to the number of cells of the target cell type (x-axis).Target cell type indicated by the color.Average accuracies over three replicates (i.e., training runs) are shown.
. The model processes input images through consecutive Inception modules, where the spatial dimensions are first contracted toward the model bottleneck (Figure 6, contracting path), and then expanded toward the output image (Figure 6, expanding path).Each subsequent Inception module in the Dee-pIFC model contained two times the number of filters contained by the preceding module, up to 512 filters.In addition, there are skip connections which connect the layers of the same spatial dimension in the contracting and expanding path.Max-pooling layers are used to reduce the input dimensionality before the model bottleneck, and then upsampling layers restore the original dimensionality to generate the fluorescent image.In addition to the fluorescent image, a set of 128 features can be extracted for each cell from the model's bottleneck layer.For UMAP visualization, features for all marker channels are concatenated to form feature files in the shape of n_cells Â 128, so all markers can be visualized in the same plot.Each DeepIFC model generates a single fluorescent image for each instance in the data.Thus, a separate DeepIFC model is trained for each fluorescent channel in the input data.DeepIFC converts each input CIF image to Hierarchical Data Format (HDF) using the Cifconvert tool [8] and extends images to 128 Â 128 size by padding edges with zero values.To normalize backgrounds in fluorescent images, each image Y is transformed with

1
), possibly due to an error in freezing and unfreezing the sample.Images were acquired at 60Â magnification with low flow rate/high sensitivity ($40 mm/s, core 7 μm), pixel size 0.33 Â 0.33 μm 2 and depth of field 2.5 μm.During the experiment, excitation lasers 405 nm (intensity 120 mW), 488 nm (intensity 145 mW), 642 nm (intensity 150 mW) were used.Fluorescent signals were gained using channels Ch02-Ch05, Ch08, Ch10 and Ch11.Channels Ch01 and Ch09 were used for brightfield (BF) images and channel 12 for the side scatter signal (SSC) from the 785 nm darkfield laser.Examples of images obtained from IDEAS ® are shown in Supplementary Figure 1.
Single color controls were used for compensation.Isotype controls and unlabelled cells were used to determine the auto fluorescence and non-specific signal.The integrated software INSPIRE ® (Luminex/ DiaSorin Group) was used for data collection.Both uncompensated and compensated images were created from the experiments with the IDEAS ® software (version 6.2).Compensated images were created by applying a compensation matrix to uncompensated image data in IDEAS ® and used in training and evaluating the DeepIFC models.In addition to images, also numerical cell features collected during the experiment (e.g., intensity of fluorescence) were extracted from IDEAS ® .We performed a gating analysis in IDEAS ® software, where positive events for each surface marker were gated based on the intensity F I G U R E 6 DeepIFC model architecture is based on Inception U-Net [35].The model inputs two brightfield and one darkfield images, and generates a predicted fluorescent image.The model consists of contracting (encoding) and expanding (decoding) paths, with a bottleneck layer in the middle outputting single-cell features.Number of filters, and input dimensionality are indicated under each module.Bottom right corner: architecture of an Inception module [36], consisting of convolutional and max-pooling layers.[Color figure can be viewed at wileyonlinelibrary.com]