Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Apr 24.
Published in final edited form as: Exp Hematol. 2025 Apr 24;147:104786. doi: 10.1016/j.exphem.2025.104786

C-COUNT: a Convolutional Neural Network-Based Tool for Automated Scoring of Erythroid Colonies

Rui Li 1, Ashley Winward 1, Logan R Lalonde 1, Daniel Hidalgo 1, John P Sardella 1, Yung Hwang 1, Aishwarya Swaminathan 1, Sean Thackeray 1, Kai Hu 1, Lihua Julie Zhu 1,*, Merav Socolovsky 1,*
PMCID: PMC12261521  NIHMSID: NIHMS2082880  PMID: 40287006

Abstract

In spite of advances in flow cytometry and single-cell transcriptomics, colony formation assays remain an essential component in the evaluation of erythroid and hematopoietic progenitors. These assays provide functional information on progenitor differentiation and proliferative potential, making them a mainstay of hematology research and clinical diagnosis. However, the utility of colony formation assays is limited by the time-consuming and error-prone manual counting of colonies, which is also prone to bias and inconsistency. Here we present ‘C-COUNT’, a convolutional neural network-based tool that scores the standard CFU-e assay by reliably identifying CFU-e colonies from images collected by automated microscopy, and outputs both their number and size. We tested the performance of C-COUNT against three experienced scientists and find that it is equivalent or better in reliably identifying CFU-e colonies on plates that also contain myeloid colonies and other cell aggregates. We further evaluated its performance in the response of CFU-e progenitors to increasing erythropoietin concentrations and to a spectrum of genotoxic agents. We provide the C-COUNT code, a Docker image, a trained model, and training dataset to facilitate its download, usage, and model refinement in other laboratories. The C-COUNT tool transforms the traditional CFU-e colony- formation assay into a rigorous and efficient assay with potential applications in high-throughput screens for novel erythropoietic factors and therapeutic agents.

Introduction

Erythroid and other hematopoietic progenitors were first defined functionally, based on the colonies they form in semi-solid medium [1–6]. Colony-formation assays (CFA) have since been joined and arguably superseded by powerful prospective phenotypic approaches that identify and define progenitors using flow-cytometric analysis of cell-surface markers [7–14], and more recently, by single-cell transcriptomic approaches [15–17]. Nevertheless, progenitor identification using these latter approaches is correlative, and requires functional confirmation. Unlike colony-formation assays, prospective phenotypic assays are unable to directly measure the differentiation and proliferative potential of progenitors. Therefore, colony formation assays have remained an essential component of progenitor cell biology and a mainstay of hematology research.

Colony formation assays were the first to demonstrate the biological spectrum of hematopoietic growth factors [18–21]. Today, they provide a powerful ‘low-tech’ tool for in vitro analysis of novel hematopoietic gene modulators and drugs [15, 22] and also play a role in clinical diagnosis, for example, in aplastic anemia, myelodysplastic syndromes, myeloproliferative disorders and cancer [23–26]. Unlike liquid cultures, which inevitably contain mixed progenitor populations, colony assays do not require expensive reagents such as antibodies to accurately capture progenitor identities. However, their large-scale use in drug or genetic screens is limited since manual colony identification and counting are slow and subject to human error and bias.

Although automated and machine-learning-based assays have been developed in recent years for counting colonies in clonogenic assays and bacterial cultures [27–34], to our knowledge, there is no dedicated and reliable tool for identification and quantification of colonies formed from colony-forming-unit-erythroid (CFU-e) progenitors. Such a tool needs to distinguish CFU-e colonies from other types of hematopoietic colonies on the same culture plates, and from non-colony cell aggregates. A commercial tool tested in the authors’ laboratory (StemVision, Stem cell technologies) did not distinguish CFU-e from other colony types. We therefore developed ‘C-COUNT’, a convolutional neural network (CNN)-based tool that makes use of recent advances in both automated microscopy and in machine learning. Machine learning refers to a collection of artificial intelligence techniques that enable systems to make inferences, predictions, and decisions based on data, whether or not explicitly programmed [35]. Powerful machine learning models are increasingly being applied in biology to analyze large and complex datasets [36–41]. Deep learning is a specialized branch of machine learning that uses neural networks with multiple layers (often referred to as deep neural networks) to model and learn from large amounts of data. These networks process and transmit information through interconnected layers, enabling the model to automatically learn intricate and hierarchical representations directly from raw data. Common deep learning architectures include Feedforward Neural Networks (FNNs), which process data from input to output in a single direction, Recurrent Neural Networks (RNNs) for sequential data like time series and natural language processing, Transformers, which are widely used in natural language processing tasks such as translation and large language models like ChatGPT, and Convolutional Neural Networks (CNNs), on which C-COUNT is based. CNNs are loosely modeled on the visual cortex [42, 43] and are particularly useful in image processing.

When evaluating C-COUNT, we found that its performance when identifying and counting CFU-e colonies is highly correlated with that of trained scientists, but that, unlike scientists, it is consistent over time. In addition to counting, C-COUNT also measures colony size. As examples of its utility, we used C-COUNT to assess the response of CFU-e progenitors to increasing concentrations of Epo, where it successfully reproduced known changes in colony number and size. We also used C-COUNT to determine previously unknown CFU-e sensitivities to a number of genotoxic agents. We provide the C-COUNT tool, along with the trained model and training dataset (https://github.com/radio1988/C-COUNT) to enable the application and customization of C-COUNT in other laboratories.

Materials and Methods

CFU-e colony-formation assays and Dab staining

Fetal liver (FL) cells (200,000 cells/ 36 mm plate) or bone-marrow (BM) cells (500,000 cells per 36 mm plate) were resuspended in medium containing 1% methylcellulose (Sigma Aldritch), 6% bovine serum albumin (BSA, Sigma Aldritch ), 5% fetal calf serum (FCS, Hyclone ) and 1mg/ml human transferrin (Sigma Aldritch). Added growth factors were Epo (Amgen) at the indicated concentrations and SCF (100 ng/ml). After 72 hours of culture, colonies were stained at room temperature with diaminobenzidine (Dab, Sigma Aldritch, 1mg/ ml in phosphate-buffered saline with 0.03% H2O2). One milliliter of the Dab solution was carefully layered on top of the methylcellulose medium for up to 30 minutes. Plates were kept at 4 °C until imaged.

Analysis of genotoxic agents

Freshly harvested FL cells were resuspended in liquid medium containing Epo (0.5 U/ml) for 60 min. They were then either irradiated with ultraviolet (UV) C light (254nm, Spectrolinker™ UV Crosslinker & Sanitizing Cabinet XL - 1500) or treated with the following genotoxic agents: hydroxyurea (Millipore Sigma, treated for 1, 2 or 4 hours), paraformaldehyde (PFA, ThermoFisher Scientific, treatment for 90 minutes) or bleomycin (Millipore Sigma, treatment for 90 min). Cells were then washed free from the genotoxic agents and plated in a methylcellulose medium for CFA as described above.

Imaging

Plates were imaged with a Zeiss Axio Observer inverted microscope with automated scanning and autofocus. Eight percent of the plate’s area was scanned in 5.5 min at a resolution of 0.5 μm per pixel. All acquired images were in grayscale, which reduced file size and processing time. Following acquisition, groups of 16 raw images in the czi format were ‘stitched’ to generate a larger image that covers an entire scanned ‘scene’. Four to eight scenes were scanned per plate.

Image preprocessing and blob detection

Stitched scene images often suffer from uneven illumination across tiles, which could affect the visibility and consistency of objects. To address this, we applied Contrast Limited Adaptive Histogram Equalization (CLAHE) [44], a method that adjusts local contrast to reduce exposure differences across the image. This process improved the uniformity, detail, and balance of objects in the stitched images (Fig. 2A).

Figure 2. Blob Detection and classification.

Figure 2

Open in a new tab

A The scene images were ‘equalized’ with Contrast Limited Adaptive Histogram Equalization (CLAHE) and then subjected to blob detection by convolving with the Laplacian of Gaussian kernel. Each individual blob is marked with a yellow circle and saved in the center of 160 × 160 pixel cropped image. The image resolution is then reduced to 80 ×80 pixels prior to further analysis.

B A portion of a stitched scene image following C-COUNT processing (see Supplementary Figure 1 for the full image). Following classification, C-COUNT outputs the stitched scene images at their original resolution, with all detected blobs marked with either a red circle (CFU-e colony) or a blue circle (non-colony). Shown here are instances where C-COUNT successfully distinguishes CFU-e from non-colony cell clusters and debris.

The equalized images were then analyzed using the Laplacian of Gaussian (LoG) method [45] to allow ‘blob’ detection. This approach repeatedly smooths the image with increasing levels of blurring and identifies regions where the intensity changes sharply, marking the presence of blob-like objects. For each detected blob, the LoG method estimates both its position and radius. Individual blobs were saved in the center of full resolution 160 × 160 pixel cropped images. The resolution of these images was then reduced to 80 × 80 pixels for all downstream analysis. Objects within the cropped image but outside the blob radius (as might happen if there were two adjacent objects) were masked to avoid interference.

Model architecture and training

We implemented a modified LeNet convolutional neural network (CNN) [46] using the open-source Keras library for model construction and TensorFlow’s Keras module for model training. The input to the model consisted of cropped and masked 80 × 80 pixel blob images. The model includes three convolutional layers, each followed by a max pooling layer, two fully connected (dense) layers, and a softmax output layer. The Rectified Linear Unit (ReLU) activation function was used in all convolutional and dense layers to facilitate training, while the softmax function in the output layer allows for easy extension to multi-class classification, such as counting different types of colonies on the same plate. To reduce overfitting and improve model generalization without resorting to model ensembles, we applied a 50% dropout rate to both convolutional and dense layers [47].

The blob dataset from CFU-e assays is highly imbalanced, containing only 1 positive blob (identified as CFU-e) for every 323 negative blobs (non-CFU-e). To address this imbalance and improve model generalization, we applied data augmentation using an ImageDataGenerator. We specifically augmented the positive blobs to match the number of negative blobs. Since we already had sufficient negative blobs, no augmentation was applied to them, except in cases where they were heavily downsampled for the saturation analysis (Fig. 3B). In addition to balancing the dataset, data augmentation helped make full use of the limited positive blob images by generating diverse variations and reducing the risk of overfitting. The augmentation techniques included random rotations (up to 90 degrees), shear transformations (up to 5%), slight zooming (up to 1%), and random width and height shifts (up to 3%). We also applied random horizontal and vertical flipping.

Figure 3. CNN architecture and training.

Figure 3

Open in a new tab

A C-COUNT model architecture, showing the LeNet CNN, with input, output and hidden layers, and their neuron number and dimensions. conv, convolution; pool, max-pooling.

B Model training (‘1’).C-COUNT The original training datasets were divided into 70% training set, 15% validation set and 15% test set. During each cycle or ‘epoch’, the model outputs classifications based on the input data, which are initially random. These predicted classifications are then compared to the classifications made by the scientist trainers. Any differences are quantified through a cross-entropy loss function. The weights of the model’s neural connections are adjusted incrementally after each epoch, so as to improve its performance, reflected by a lower cost function value after the next iteration. When the cost function of the validation set stops improving, training is terminated, to avoid overfitting the model to the training set. Following training, model performance is evaluated against an independent test dataset (‘2’).

C Size calculation was performed for positive ‘blobs’ (CFU-e colonies) using the Yen method (see text for details).

D Saturation analysis. Model performance was measured using Matthew’s correlation coefficient (MCC). The model was trained multiple times. Each training session used a randomly-selected subset of the original training dataset. Each grey marker represents a trained C-COUNT algorithm, and each red marker corresponds to the average MCC (± standard error) after training with a specific fraction of the original dataset.

E Flow-chart illustrating the C-COUNT workflow. Input stitched images are processed by the algorithm, which outputs spreadsheets containing colony count data and colony area data. C-COUNT also outputs all the input images with marked positive and negative classifications (see supplementary Figure 1).

During training, categorical cross-entropy was used as the loss function. The Adaptive Moment Estimation (Adam) optimizer was employed for efficient gradient descent optimization. To enhance model performance and training stability, a mini-batch size of 256 cropped images was used, and the learning rate was set to 0.001. To further prevent overfitting, early stopping was implemented based on validation loss.

Colony size calculation

Colony sizes were calculated for all colonies classified as positive as follows: For each cropped blob image, the region outside the blob’s radius was masked and excluded from the calculation. Within the radius, the Yen [48] method was used to distinguish the foreground (colony) from the background based on pixel intensity levels. The area of the foreground, in pixels, was then calculated to represent colony size. This pixel-based size was converted to square microns using a conversion factor based on our imaging setup, which was validated using sized beads (1 pixel = 2.16 microns).

Downstream Data analysis and Figure generation

Downstream data analysis and figure generation were performed using Python 3.9 and Jupyter Lab 4.4, along with the libraries matplotlib, seaborn, and statsmodels.

Results

Optimization of the CFU-e assay and preparation of training datasets

Mouse CFU-e progenitors are found in the fetal liver (FL) at mid-gestation and in adult bone-marrow (BM) and spleen. They give rise to closely-packed small spherical colonies containing 16 to 32 erythroblasts or red cells within 48 to 72 hours of plating in semi-solid medium such as methylcellulose in the presence of erythropoietin (Epo). CFU-e colony-formation assays (CFAs) are inherently variable for a number of reasons, including variability in the state of cells harvested from primary tissue, variability in the biological activity of media components, the difficulty of delivering an accurate volume of semi-solid medium into replicate plates, and the relatively low progenitor cell densities required for successful colony counting in these assays. Here we first established a standard protocol that attempts to minimize these sources of variability. We experimented with media containing a range of methylcellulose densities, finding that in 1% methylcellulose CFU-e colonies are largely within the same focal plane during imaging. We tested a number of optical-bottom plates designed for imaging of cultured cells, but found no advantage over routinely-used tissue culture plastic. Testing of plates with hydrophobic plastic that minimizes the formation of a meniscus along the plate perimeter suggested that the expense is not justified. Instead, we imaged equivalent areas at the same distance from the perimeter in all plates. We next tested media components, and determined that 5% FCS supplemented with 6% BSA give low background of colony-formation in the absence of Epo, and good yields of Epo-dependent CFU-e colonies (Fig. 1A). We also optimized plated cell density. High densities appear to somewhat increase the number of CFU-e colonies formed per million cells plated, but also make counting more challenging as colonies run into each other (Methods, Fig. 1B). Finally, we devised a method for correcting inaccuracies in aliquoting precise volumes of methylcellulose. Methylcellulose medium containing uniformly suspended progenitors is viscous and is routinely dispensed into each plate with a syringe and needle. Some variability in the amount dispensed to each plate is inevitable, however. Therefore, we measured the actual volume dispensed by weighing each plate, and applied a correction factor based on this weight to the number of colonies counted in each plate.

Figure 1. Standardization of the CFU-e colony formation assay.

Figure 1

Open in a new tab

A Effect of FCS on background (Epo=0) and Epo- dependent CFU-e colony formation

B Effect of cell densities on CFU-e colony formation. The top panel documents the number of colonies formed per plate, which increases with increasing number of cells plated. The bottom panel calculates the number of CFU-e colonies formed per million cells plated.

C Colonies were stained in situ with Dab, and kept horizontal on ice until imaged

D Time-lapse imaging during Dab staining. Top panel shows a CFU-e colony, bottom panel a nonCFU-e colony.

E A diagram illustrating the circular plate and the four scenes scanned and imaged by the Zeiss Axio Observer inverted microscope. The inset shows the 16 smaller images that are stitched together to obtain the complete scene image.

The only essential growth factor for CFU-e colony formation is Epo. Other growth factors such as Stem Cell Factor (SCF) improve CFU-e survival and increase colony number [49], but also support the formation of myeloid colonies. We trained C-COUNT on images of CFU-e assay plates that contained either no growth factors (‘background’ plates with virtually no CFU-e colonies), Epo alone (giving largely CFU-e colonies and few other colony types), or both Epo and SCF (where there are both CFU-e and myeloid colonies). The training datasets contained all of the CFU-e colony images from two CFAs using adult mouse BM and mouse FL at mid-gestation (Table 1). To increase the number of CFU-e colonies in our training dataset, we also included CFAs done on sorted Ter119- FL cells, which are enriched in CFU-e activity. We compared the performance of preliminary C-COUNT algorithms trained on images derived exclusively from either FL or BM and observed no difference in their performance when tested on CFU-e colony images from either tissue. For simplicity, therefore, the final version of the C-COUNT tool was trained, validated and tested using datasets from both FL and BM (Table 1).

Table 1:

Datasets in the training, validation and testing of c-count

Dataset Tissue Number of labelers Analysis type EPO (U/ml) SCF (ng/ml)
1. Model training
Training (70%) Validation (15%) Testing (15%)
Negative Positive Negative Positive Negative Positive
A FL, Ter119- 1 15256 419 3267 92 3247 113 0, 0.5 100
F BM 3 233702 350 50086 68 50093 62 DRC*, 0 to 1
A and F 248958 769 53353 160 53340 175
2. Performance evaluation: manual counting vs. C-COUNT
B FL 3 DRC, 0 to 1
E BM 3 DRC, 0 to 1
3. Consistency analysis over time: manual counting vs. C-COUNT
FL2 FL 1 Comparing scientists and c-count 0.5
4. Analysis using the trained C-COUNT tool
FL DRC with the DNA damaging agents: paraformaldehyde, bleomycin, ultraviolet irradiation, hydroxyurea 0.5 100
Open in a new tab
*

DRC, dose-response curve

Diaminobenzidine staining and imaging of colonies

Hemoglobin has mild peroxidase activity, which may be used to visualize hemoglobinized cells in mature CFU-e colonies. After 72 hours of culture, colonies were stained with the peroxidase substrate diaminobenzidine (Dab), increasing confidence in their erythroid identity [50] (Fig. 1C-D). In addition, the dark Dab stain increases the contrast between CFU-e colonies and surrounding medium, improves the autofocus function of automated microscopes and their prominence in the acquired images. To determine the optimal staining time, we imaged the same fields at 8 min intervals during staining at room temperature. We found that CFU-e colonies acquire the stain rapidly, becoming uniformly brown by 16 to 32 minutes; by contrast, non-erythroid cells remain largely pale during this period (Fig. 1D; note that the non-CFU-e cell cluster is composed largely of non-erythroid cells but also contains a small number of associated red cells). Because the viscose medium is slow to correct if tilted, plates were kept on ice perfectly horizontal prior to imaging. Stained colonies were imaged using an automated inverted microscope, which scanned 4–8 identical regions, or ‘scenes’, scattered just inside the perimeter of each plate, totaling 8 to 16% of the plate’s area (Fig. 1E).

Image pre-processing and labeling

To minimize variability between images, the brightness and contrast of each stitched scene image was digitally equalized [44] (Fig. 2A). We then applied a ‘blob detection’ algorithm to the images, by identifying regions that differ from their surrounding in brightness or other properties (Fig. 2B). Blobs were classified manually by trained scientists as either positive (CFU-e colony present) or negative (non- colony ‘blob’). In preliminary experiments, manual classification was performed on cropped images each containing a single blob. This initial labeled dataset was used for preliminary training of C-COUNT. However, we found that labeling of single blobs without the context of the field of view was challenging. We therefore developed a second manual labeling approach. Here the labeler sees the entire scene image, containing multiple blobs. Each blob is circled by either a blue or a red circle, denoting initial classification by the preliminary C-COUNT algorithm as negative (non-CFU-e) or positive (CFU-e), respectively. The labeler marked all positive colonies, thereby correcting any misclassifications. The preliminary C-COUNT algorithm generated a large number of false positives. To overcome this, we increased the number of negative blobs in the training datasets, by using images from plates that had no Epo, and assuming all blobs on these plates were negative; this successfully reduced the number of false positives.

Model architecture and training

The cropped and masked blob images were used as input for a LeNet CNN [46] (Fig. 3A). The CNN’s layers include three convolutional layers, max-pooling layers, densely connected layers, and dropout layers (not shown). Max-pooling layers condense the data and reduce the number of parameters by retaining the highest value from small regions of the previous layer. The dropout layers randomly deactivate a fraction of the network’s connections during training, which helps prevent overfitting of the model to the training dataset.

Model fitting is an iterative process (Fig. 3B). In this study, we used 70% of the combined FL and BM dataset for training, 15% for validation, and 15% for model testing (Table 1). During each training cycle or ‘epoch’, the input training set is used by the model to generate predicted classifications. These are then compared to the actual ‘target labels’ for the data. A comparison is also made between predicted classifications and target labels in an independent validation set. The difference between the predicted and actual labels is quantified through a ‘cost function’. Training continues until the cost function no longer shows improvement, at which point the process is terminated. Finally, the model is evaluated using an independent test dataset.

To assess visually the quality of C-COUNT classifications, it outputs scene images where each blob is circled and positive blobs are indicated (Fig. 2B, supplementary Fig 1).

In addition to counting the number of colonies classified as positives, C-COUNT also measures their area (Fig. 3C.).

Saturation Analysis

To assess whether a sufficiently large dataset was used for training, and to determine the minimal training dataset size that would be required to set up C-COUNT in another laboratory, we carried out saturation analysis (Fig. 3D). This analysis evaluates the performance of the C-COUNT tool after training on a randomly -selected subset of the original dataset. To account for the inherent variability in model training, the process was repeated five times for each dataset size. Model performance was evaluated using the Matthews Correlation Coefficient (MCC), which incorporates all four possible classification outcomes—true positives, false positives, true negatives and false negatives—via a confusion matrix. Optimal training performance was achieved with 12.5% of the full-size dataset, which included ~ 100 positive colonies and ~ 30,000 negative blobs. It is important to note that many of the negative blobs used for training were sourced from ‘background’ plates grown in the absence of Epo. These blobs were treated as negative during training without the need for manual labeling.

Input and output of the trained C-COUNT tool

During a routine experiment, the inputs to the C-COUNT tool are stitched scene images each comprising approximately 2% of the area of a 35 mm colony plate (Supplementary Fig. 1, Fig. 1E, Fig. 3E). We use 4 to 8 such scenes per plate, and 2 to 4 replicate plates for each test condition in our experiments. The C-COUNT algorithm then processes the images, including blob detection, cropping, and classification. It generates two spreadsheets: one containing the count data for each scene and the other detailing the area for each positive colony in each scene (Fig. 3C, E). Additionally, it produces images of all the scenes, with each detected blob circled in either red (CFU-e colony) or blue (non-colony). For an example of a scene image following processing, see Supplementary Fig.1 and Fig. 2B.

Evaluation of C-COUNT performance in Epo dose-response analysis

We evaluated the C-COUNT tool in two experiments that tested the effect of Epo concentration on either FL or BM CFU-e colony growth. The datasets generated by these experiments were independent of the original training datasets (Table 1). Additionally, images from these two experiments were manually analyzed by three scientists, each of whom labeled all of the images independently. This approach allowed us to assess inter-scientist variability in manual colony labeling, and to compare it with the C-COUNT results. We found considerable variation in the number of blobs identified as ‘CFU-e’ by scientists; however, the results by any given scientist were strongly correlated with those of the others. Similarly, there was a strong correlation between C-COUNT’s results and those of each of the scientists. Thus, for the FL dataset, Pearson’s correlation coefficients between plate colony-counts by C-COUNT and by each of the scientists were 0.92, 0.94 and 0.96, whereas correlation coefficients amongst the scientists were 0.86, 0.90 and 0.93, Fig. 4A. Similar trends were observed for the BM dataset, Supplementary Fig. 2A). The Epo dose-response curve (DRC) generated by C-COUNT matched closely that generated by the scientists (Fig. 4B). Further, the scene-level correlation coefficients between C-COUNT and each of the scientists across the two experiments were significantly better than the correlation coefficients between the three scientists (p<0.001, linear mixed-effects model, Supplementary Fig.2B ). This suggests that C-COUNT’s reproducibility is either on par with or exceeds that of manual counting by the scientists.

Figure 4. C-COUNT performance evaluation.

Figure 4

Open in a new tab

A Analysis of an Epo dose-response (DRC) experiment using FL cells (dataset B, Table 1), scored by C-COUNT, and manually by three trained scientists. Shown are correlations between C-COUNT ‘s classification and each manual classification (‘m1’, ‘m2’, ‘m3). Also shown are correlations between the manual classifications. Blue markers are colony number in each scene. Red markers are colony number per plate. Pearson’s correlation coefficients are indicated on each scatter plot. Additionally, colony number distributions per plate, as scored by each manual count or by C-COUNT, are shown with the total number of CFU-e colonies detected . See also supplementary Figure 2 for equivalent analysis on a BM dataset (dataset E).

B Epo DRCs on FL or BM progenitors, scored manually by each of three scientist, and by the trained C-COUNT algorithm. The same datasets as in ‘A’ and in supplementary Figure 2.

C Changes in colony-area distribution in response to increasing Epo concentration. FL cell dataset as in ‘A’ and ‘B’.

Colony area measurements by C-COUNT showed that higher levels of Epo support the formation of larger CFU-e colonies (Fig. 4C). This reproduces our recent findings using manual colony size measurements on a much smaller colony dataset (20 colonies) using ImageJ [51]. Hence, by virtue of its ability to measure colony areas on the entire colony dataset of several hundred colonies, C-COUNT increases confidence in this finding.

Consistency of manual vs. C-COUNT over time

To further evaluate C-COUNT and compare it with manual counting, we asked how manual counting by the same scientist might vary over time. We used an independent set of 1500 ‘blobs’ (Table 1) that were classified manually twice, 18 days apart, by the same scientist. As anticipated, we observed strong, though not perfect, correlations between the first and second manual classifications (MCC of 0.73, Fig. 5A, supplementary Table 1). C-COUNT’s performance matched well with the manual classification (MCC= 0.80, Fig. 5B). Since C-COUNT’s classification of a given image is deterministic, the two runs of C-COUNT on the same dataset yielded perfectly correlated results (MCC=1, Fig. 5C). These results show that C-COUNT’s performance is well within the expected range for trained scientists, and further, that one of its advantages is that of consistency over time.

Figure 5. Consistency analysis.

Figure 5

Open in a new tab

A Manual classifications of the same 1500 ‘blobs’ (m1), repeated 18 days apart. MCC measures the consistency between both classifications.

B Correlation between the manual classification and C-COUNT

C Perfect correlation (MCC = 1) between classifications in the first and second runs of the same dataset by C-COUNT. Same dataset as in A and B.

Use of the C-COUNT tool in analysis of CFU-e sensitivity to genotoxic agents

As an example of the potential applications of C-COUNT, we examined the sensitivity of CFU-e progenitors to a spectrum of distinct genotoxic agents: ultraviolet light (UV), hydroxyurea (HU), paraformaldehyde (PFA) and bleomycin. Each of these agents causes distinct DNA lesions that elicit a unique set of DNA damage response pathways. We harvested fresh FL CFU-e and allowed them to recover for 60 minutes at 37°C in Epo- containing medium. We then subjected the cells to a brief period of genotoxic stress; their ability to recover was measured using CFU-e CFA. The cells were subjected to UV light, or alternatively, a genotoxic drug was added to the culture medium for 1 to 4 hours (Fig. 6). Cells were then washed and transferred to methylcellulose medium in the presence of Epo and SCF for CFA. Plates were imaged at 72 hours, and images were analyzed using C-COUNT. In all cases, colony number decreased with increasing genotoxic stress (Fig. 6; IC50 was determined for genotoxic agents with a full dose-response curve). Surprisingly, colony areas were not affected by genotoxic treatment with experimental design we employed, suggesting that CFU-e progenitors were either irreversibly damaged, failing to give rise to colonies, or else were able to recover fully after the genotoxic agent was removed (Fig. 6).

Figure 6. CFU-e sensitivity to genotoxic agents: C-COUNT analysis.

Figure 6

Open in a new tab

A, B Colony number (A) and colony diameter distributions (B) for CFU-e colonies in CFA following irradiation of FL cells with a range of UV light energies. IC50 was determined by fitting the colony number data with the Hill equation. Violin plots show individual colonies (dots), medians and upper and lower quartiles. Diameters were calculated from colony area in pixels, using the measured resolution of 2.16 pixels per micron.

C- F As in A,B, but FL cells were treated with either paraformaldehyde (for 90 min, C, D), hydroxyurea (for 2 or 4 hours, E, F) or with bleomycin (for 90 min, G, H). Following treatment, cells were washed and then plated for CFA.

Discussion

Here we present an open-source tool for classifying and counting CFU-e colonies in a CFA using hematopoietic tissue such as whole BM or FL. The C-COUNT tool identifies CFU-e colonies, distinguishing them from myeloid colonies and from non-colony cell aggregates or debris. It outputs both colony number and colony size as well as images where the CFU-e colonies are marked. The C-COUNT tool may be used for large-scale CFU-e experiments and for drug screens, saving labor and avoiding bias in scoring colonies.

In evaluation experiments, C-COUNT’s performance correlated closely with manual counts by the three scientists that trained the tool. The correlation in performance between C-COUNT and each of the scientists was significantly higher than the correlation between scientists, although all correlations were high (ranging from 0.88 to 0.97). As may be expected, we found that even a trained scientist is not perfectly consistent in their scoring of colonies over time. In this respect, it is clearly advantageous to use a tool that applies a consistent and deterministic classification algorithm across multiple experiments. Finally, we show an example of C-COUNT’s performance assessing the effect of genotoxic drugs on CFU-e number and colony area distribution.

Although we started with relatively high resolution images to aid accurate labeling by the scientists when generating the training datasets, the input to the C-COUNT tool during training and subsequent use was of images with a quarter of the initial resolution. It is therefore possible to use the tool with lower resolution images, which would speed up image acquisition. Alternatively, larger areas of the plate could be scanned at lower resolution during automated microscopy. Users of the tool can adjust these parameters based on available microscopy.

Future work may address the development of CNN tools that identify a wider range of hematopoietic colonies, perhaps using color microscopy in combination with immunohistochemistry. The application of CNN-based image analysis to hematopoietic colony formation assays is in its infancy but has vast potential in both fundamental and translational research.

Supplementary Material

Supp.File

Supplementary Figure 1 Stitched image with classifications

An example of a stitched image of a scene, following classification. Each ‘blob’ is marked with either a red circle (CFU-e colony) or a blue circle (negative for CFU-e classification). The image is stitched from 16 individual images which together make up 1.9% of the plate area.

Supplementary Figure 2 C-COUNT performance evaluation: correlations with BM CFU-e classifications

A BM dataset (dataset E, Table 1) from an Epo dose-response experiment that was scored both manually and by C-COUNT. Shown are correlations between the classification by C-COUNT and each manual classification (‘m1’, ‘m2’, ‘m3), as well as correlations between the three manual classifications. Blue markers are colony number in each scene. Red markers are colony number per plate. Numbers on the scatter plots are Pearson’s correlation coefficients. Also shown are colony number distributions in each plate, scored by each scientist or by C-COUNT, with the total number of CFU-e colonies detected noted. See also Figure 4A for equivalent analysis on a FL dataset (dataset B).

B Linear mixed-effects model (LMM) analysis comparing the correlations between manual counts (m1, m2, m3 in panel A and in Figure 4A, datasets B and E) with the correlations between C-COUNT and each of the same manual counts. The boxplots show the distribution of the Pearson correlation values for each comparison. The LMM was applied using the type of comparison (between manual counts, or between C-COUNT and manual counts) as a fixed effect, and the dataset (B or E) as a random effect.

Highlights.

  • C-COUNT is a neural-network-based tool for analyzing CFU-e colonies

  • The input to C-COUNT are automated microscopy images of CFU-e assay plates

  • C-COUNT distinguishes CFU-e from other colonies and outputs their number and size

  • C-COUNT’s performance is highly correlated with that of trained scientists

  • C-COUNT has potential applications in large-scale CFU-e assays and drug screens

Acknowledgement

This work was funded by NIH/NIDDK grants R01DK120639, R01DK130498 and R01DK136321. Sorted cells were obtained using BD FACSAria Fusion Cell Sorter purchased with NIH equipment grant 1S10OD028576.

Footnotes

Competing interest statement

The authors have no competing interests.

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Data availability

All training data, trained C-COUNT weight h5 files, example images and labels are described on a GitHub page (https://github.com/radio1988/C-COUNT). Instructions are also provided to apply C-COUNT or to re-train it using new images. We also provided a Q & A page for C-COUNT usage related questions (https://github.com/radio1988/C-COUNT/issues).

References

  • [1].Metcalf D. Regulation of hemopoiesis. Nouv Rev Fr Hematol. 1978;20:521–533. [PubMed] [Google Scholar]
  • [2].Schrader JW. Bone marrow differentiation in vitro. Crit Rev Immunol. 1983;4:197–277. [PubMed] [Google Scholar]
  • [3].Suda T, Suda J, Ogawa M. Disparate differentiation in mouse hemopoietic colonies derived from paired progenitors. Proc Natl Acad Sci USA. 1984;81:2520–2524. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [4].Richert-Boe KE, Bagby GC Jr. In vitro hematopoiesis in myelodysplasia: liquid and soft-gel culture studies. Hematol Oncol Clin North Am. 1992;6:543–556. [PubMed] [Google Scholar]
  • [5].Ogawa M. Differentiation and proliferation of hematopoietic stem cells. Blood. 1993;81:2844–2853. [PubMed] [Google Scholar]
  • [6].Socolovsky M, Constantinescu SN, Bergelson S, Sirotkin A, Lodish HF. Cytokines in hematopoiesis: specificity and redundancy in receptor function. Adv Protein Chem. 1998;52:141–198. [DOI] [PubMed] [Google Scholar]
  • [7].Morrison SJ, Weissman IL. The long-term repopulating subset of hematopoietic stem cells is deterministic and isolatable by phenotype. Immunity. 1994;1:661–673. [DOI] [PubMed] [Google Scholar]
  • [8].Kondo M, Weissman IL, Akashi K. Identification of clonogenic common lymphoid progenitors in mouse bone marrow. Cell. 1997;91:661–672. [DOI] [PubMed] [Google Scholar]
  • [9].Akashi K, Traver D, Miyamoto T, Weissman IL. A clonogenic common myeloid progenitor that gives rise to all myeloid lineages. Nature. 2000;404:193–197. [DOI] [PubMed] [Google Scholar]
  • [10].Forsberg EC, Serwold T, Kogan S, Weissman IL, Passegue E. New evidence supporting megakaryocyte-erythrocyte potential of flk2/flt3+ multipotent hematopoietic progenitors. Cell. 2006;126:415–426. [DOI] [PubMed] [Google Scholar]
  • [11].Weissman IL, Shizuru JA. The origins of the identification and isolation of hematopoietic stem cells, and their capability to induce donor-specific transplantation tolerance and treat autoimmune diseases. Blood. 2008;112:3543–3553. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [12].Mori Y, Chen JY, Pluvinage JV, Seita J, Weissman IL. Prospective isolation of human erythroid lineage-committed progenitors. Proceedings of the National Academy of Sciences of the United States of America. 2015;112:9638–9643. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [13].Liu Y, Pop R, Sadegh C, Brugnara C, Haase VH, Socolovsky M. Suppression of Fas-FasL coexpression by erythropoietin mediates erythroblast expansion during the erythropoietic stress response in vivo. Blood. 2006;108:123–133. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [14].Swaminathan A, Hwang Y, Winward A, Socolovsky M. Identification and Isolation of Burst-Forming Unit and Colony-Forming Unit Erythroid Progenitors from Mouse Tissue by Flow Cytometry. J Vis Exp. 2022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [15].Tusi BK, Wolock SL, Weinreb C, et al. Population snapshots predict early haematopoietic and erythroid hierarchies. Nature. 2018;555:54–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [16].Velten L, Haas SF, Raffel S, et al. Human haematopoietic stem cell lineage commitment is a continuous process. Nat Cell Biol. 2017;19:271–281. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [17].Regev A, Teichmann SA, Lander ES, et al. The Human Cell Atlas. Elife. 2017;6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [18].Metcalf D. The hematopoietic colony stimulating factors. Amsterdam. New York. Oxford: Elsevier; 1984. [Google Scholar]
  • [19].Metcalf D, Begley CG, Nicola NA, Lopez AF, Williamson DJ. Effects of purified bacterially synthesised murie multi-CSF (IL-3) on hematopoiesis in normal adult mice. Blood. 1986;68:46–57. [PubMed] [Google Scholar]
  • [20].Metcalf D. Control of granunocytes and macrophages: molecular, cellular and clinical aspects. Science (New York, NY). 1991;254:529–533. [DOI] [PubMed] [Google Scholar]
  • [21].Fletcher FA, Williams DE. Recent progress in the discovery and invention of novel hematopoietic cytokines. Crit Rev Oncol Hematol. 1992;13:1–15. [DOI] [PubMed] [Google Scholar]
  • [22].Wu QC, Swaminathan A, Winward A, et al. IL-17A primes an early progenitor compartment to tune the erythropoietic feedback circuit. bioRxiv. 2024:2024.2012.2016.628761. [Google Scholar]
  • [23].Tefferi A, Spivak JL. Polycythemia vera: scientific advances and current practice. Semin Hematol. 2005;42:206–220. [DOI] [PubMed] [Google Scholar]
  • [24].Ohler L, Geissler K, Hinterberger W. Diagnostic and prognostic value of colony formation of hematopoietic progenitor cells in myeloid malignancies. Wien Klin Wochenschr. 2003;115:537–546. [DOI] [PubMed] [Google Scholar]
  • [25].Li B, Liu J, Qu S, et al. Colony-forming unit cell (CFU-C) assays at diagnosis: CFU-G/M cluster predicts overall survival in myelodysplastic syndrome patients independently of IPSS-R. Oncotarget. 2016;7:68023–68032. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [26].Franken NA, Rodermond HM, Stap J, Haveman J, van Bree C. Clonogenic assay of cells in vitro. Nat Protoc. 2006;1:2315–2319. [DOI] [PubMed] [Google Scholar]
  • [27].Zhang L. Machine learning for enumeration of cell colony forming units. Vis Comput Ind Biomed Art. 2022;5:26. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [28].Jumutc V, Suponenkovs A, Bondarenko A, Bļizņuks D, Lihachev A. Hybrid Approach to Colony-Forming Unit Counting Problem Using Multi-Loss U-Net Reformulation. Sensors (Basel). 2023;23. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [29].Guzmán C, Bagga M, Kaur A, Westermarck J, Abankwa D. ColonyArea: an ImageJ plugin to automatically quantify colony formation in clonogenic assays. PLoS One. 2014;9:e92444. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [30].Gomes NP, Frederick B, Jacobsen JR, Chapnick D, Su TT. A High Throughput Screen with a Clonogenic Endpoint to Identify Radiation Modulators of Cancer. Radiat Res. 2023;199:132–147. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [31].Choudhry P. High-Throughput Method for Automated Colony and Cell Counting by Digital Image Analysis Based on Edge Detection. PLOS ONE. 2016;11:e0148469. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [32].Carl SH, Duempelmann L, Shimada Y, Bühler M. A fully automated deep learning pipeline for high-throughput colony segmentation and classification. Biol Open. 2020;9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [33].Alsulimani A, Akhter N, Jameela F, et al. The Impact of Artificial Intelligence on Microbial Diagnosis. Microorganisms. 2024;12. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [34].Brzozowska B, Galecki M, Tartas A, Ginter J, Kazmierczak U, Lundholm L. Freeware tool for analysing numbers and sizes of cell colonies. Radiat Environ Biophys. 2019;58:109–117. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [35].DataCamp. Machine Learning for Everyone. [Google Scholar]
  • [36].van der Maaten L. Accelerating t-SNE using Tree-Based Algorithms. Journal of Machine Learning Research 2014;15:3221–3245. [Google Scholar]
  • [37].Scialdone A, Natarajan KN, Saraiva LR, et al. Computational assignment of cell-cycle stage from single-cell transcriptome data. Methods. 2015;85:54–61. [DOI] [PubMed] [Google Scholar]
  • [38].Min S, Lee B, Yoon S. Deep learning in bioinformatics. Brief Bioinform. 2017;18:851–869. [DOI] [PubMed] [Google Scholar]
  • [39].Camacho DM, Collins KM, Powers RK, Costello JC, Collins JJ. Next-Generation Machine Learning for Biological Networks. Cell. 2018;173:1581–1592. [DOI] [PubMed] [Google Scholar]
  • [40].Ching T, Himmelstein DS, Beaulieu-Jones BK, et al. Opportunities and obstacles for deep learning in biology and medicine. J R Soc Interface. 2018;15. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [41].Kusumoto D, Yuasa S. The application of convolutional neural network to stem cell biology. Inflamm Regen. 2019;39:14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [42].Ciregan D, Meier U, Schmidhuber J. Multi-column deep neural networks fo rimage classification. IEEE Xplore; 2012. [DOI] [PubMed] [Google Scholar]
  • [43].Liao Q, Poggio T. Bridging the Gaps Between Residual Learning, Recurrent Neural Networks and Visual Cortex. arXiv. 2016;arXiv:1604.03640 [cs.LG]. [Google Scholar]
  • [44].Zuiderveld K. Contrast Limited Adaptive Histograph Equalization. Graphics Gems IV: Academic Press Professional; 1994. p. 474–485. [Google Scholar]
  • [45].Sotak GEJ, Boyer KL. The laplacian-of-gaussina kernel: a formal analysis and design procedure for fast, accurate confolution and full-frame output. Computer Vision, Graphics, and Image Processing. 1989;48:147–189. [Google Scholar]
  • [46].LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-Based Learning Applied to Document Recognition. Proc of The IEEE. 1998. [Google Scholar]
  • [47].Hinton GE, Srivastava N, Krizhevsky A, Sutskever I, Salakhutdinov RR. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:12070580. 2012. [Google Scholar]
  • [48].Yen JC, Chang FJ, Chang S. A new criterion for automatic multilevel thresholding. IEEE Trans Image Process. 1995;4:370–378. [DOI] [PubMed] [Google Scholar]
  • [49].Socolovsky M, Fallon AEJ, Lodish HF. The prolactin receptor rescues EpoR −/− erythroid progenitors and replaces EpoR in a synergistic interaction with c-kit. Blood. 1998;92:1491–1496. [PubMed] [Google Scholar]
  • [50].Socolovsky M, Dusanter-Fourt I, Lodish HF. The Prolactin receptor, as well as severly truncated erythropoietin receptors support differentiation of erythroid progenitors. J Biol Chem. 1997;272:14009–14013. [DOI] [PubMed] [Google Scholar]
  • [51].Hidalgo D, Bejder J, Pop R, et al. EpoR stimulates rapid cycling and larger red cells during mouse and human erythropoiesis. Nature communications. 2021;12:7334. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supp.File

Supplementary Figure 1 Stitched image with classifications

An example of a stitched image of a scene, following classification. Each ‘blob’ is marked with either a red circle (CFU-e colony) or a blue circle (negative for CFU-e classification). The image is stitched from 16 individual images which together make up 1.9% of the plate area.

Supplementary Figure 2 C-COUNT performance evaluation: correlations with BM CFU-e classifications

A BM dataset (dataset E, Table 1) from an Epo dose-response experiment that was scored both manually and by C-COUNT. Shown are correlations between the classification by C-COUNT and each manual classification (‘m1’, ‘m2’, ‘m3), as well as correlations between the three manual classifications. Blue markers are colony number in each scene. Red markers are colony number per plate. Numbers on the scatter plots are Pearson’s correlation coefficients. Also shown are colony number distributions in each plate, scored by each scientist or by C-COUNT, with the total number of CFU-e colonies detected noted. See also Figure 4A for equivalent analysis on a FL dataset (dataset B).

B Linear mixed-effects model (LMM) analysis comparing the correlations between manual counts (m1, m2, m3 in panel A and in Figure 4A, datasets B and E) with the correlations between C-COUNT and each of the same manual counts. The boxplots show the distribution of the Pearson correlation values for each comparison. The LMM was applied using the type of comparison (between manual counts, or between C-COUNT and manual counts) as a fixed effect, and the dataset (B or E) as a random effect.

NIHMS2082880-supplement-Supp_File.pdf (1,010.1KB, pdf)

Data Availability Statement

All training data, trained C-COUNT weight h5 files, example images and labels are described on a GitHub page (https://github.com/radio1988/C-COUNT). Instructions are also provided to apply C-COUNT or to re-train it using new images. We also provided a Q & A page for C-COUNT usage related questions (https://github.com/radio1988/C-COUNT/issues).

RESOURCES