A human epithelial cell cycle dataset

Abstract

We introduce a large-scale single cell dataset for cell cycle phases. The dataset includes single cell images of 4′,6-diamidino-2-phenylindole-stained human epithelioma-2 (HEp-2) cells grown on microscopic glass slides. Greyscale images of adherent HEp-2 cells were acquired at 20x magnification by an automated, inverse microscope system. Cell images were automatically segmented with 256 × 256 pixel patch size including surrounding area. Initially, 10,000 images were manually classified by experts into 5 cell-cycle groups (interphase, prophase, metaphase, anaphase, telophase) and two additional groups (artefact, triple). Based on these initial single cell-image classifications, a convolutional neural network (CNN) model (VGG16) was trained. Subsequently, additional segmented single cell images were pre-classified by the CNN and double-blinded reviewed as well as corrected by two experts to build an algorithm-aided dataset with 100,000 images. The dataset provides a resource for development of biomedical and deep learning applications and, thus, enables training and assessment of cell-cycle detection algorithms.

Background & Summary

The detection of cell cycle-specific patterns obtained by various microscopy and imaging technologies has become central to many diagnostic applications^1–3. The recognition and classification of mitotic figures is one of the most frequent tasks in this context in various fields such as tumor or autoimmune diagnostics. Thus, the number of mitotic figures, stained generally by hematoxylin and eosin, was reported as one of the strongest prognostic factors in tumor growth assessment^4,5. Likewise, mitotic figures and their fluorescent staining by immune markers such as antinuclear antibodies (ANA) play a central role in the serological diagnosis of autoimmune diseases^6,7. This applies in particular to the use of automated computer-aided diagnosis (CAD) systems^2,8,9.

In general, the identification of mitotic cells and their classification into different mitotic phases is prone to high variability as a result of human bias, morphological variability, and different distribution of mitotic cells in tumour tissues or cell lines¹⁰. Thus, inter-rater agreement statistics revealed poor agreement between pathologists with Cohen’s kappa values ranging from 0.13 and 0.39 with respect to the identification of single mitotic cells in tumour tissue sections¹¹. Inter-rater agreement quantifies kappa statistic in accordance with Cohen to evaluate the agreement between two classifications¹².

Specifying mitotic patterns for ANA testing also remains a challenge, although the introduction of the first international consensus on ANA patterns (ICAP) (www.anapattern.org) in 2015 sought to improve harmonization of ANA testing by indirect immunofluorescence assay (IFA) using HEp-2 cells as substrate¹³. The mitotic patterns classified as anti-cell (AC)-24 to AC-28 require the identification of mitotic phases while the staining of the metaphase plate is essential for reporting patterns of the competent and expert levels such as AC-1 (homogenous), AC-2/30 (speckled with mitotic plate), AC-3 (centromere) or AC-29 (DNA topoisomerase I) (www.anapattern.org). Since the first attempts to use CAD-interpretation systems for ANA reporting, the recognition of mitotic figures and definition of mitotic phases either by mathematical pattern-recognition algorithms and recently machine learning have been in the limelight^14–17.

Thus, we created an annotated scale and intensity-independent HeLa/HEp-2 cell cycle-image dataset (https://doi.org/10.5281/zenodo.18337703) for deep learning that is based on initial CNN models for biomedical research applications, which could support ANA pattern classification and tumor-associated research¹⁸. The dataset contains images of variable exposure times showing cells with lower and higher brightness. This allows the development of intensity-independent robust models. Each cell contains surrounding objects, which makes the set scale independent. The use of a combination of HeLa and HEp-2 images is aimed to reflect that HEp-2 cells originally thought to be derived from an epidermoid carcinoma of the larynx could be contaminated with HeLa cells based on isoenzyme analysis, HeLa marker chromosome occurrence, and DNA fingerprinting as shown for the American type culture collection (ATCC®) HEp-2 cell line CCL-23™ (www.atcc.org).

Greyscale images of 4′,6-diamidino-2-phenylindole (DAPI)-stained HEp-2 cells were acquired by an automated, inverse microscope system. Single cell images were automatically segmented with a resolution of 256 × 256 pixel. Several VGG-Net architecture CNNs were employed to learn important features for cell cycle-classification. Potential mitotic figures were identified by two immunologists [D.Ro., R.Hi.] and manually classified into 5 cell-cycle groups (interphase, prophase, metaphase, anaphase, telophase) and two additional groups (artefact, triple). Subsequently, additional images were classified by a deep learning-based pipeline (VGG16). VGG16 is a foundational CNN architecture designed for image classification tasks. It is renowned for its uniform, layer-by-layer structure, relying heavily on consecutive 3 × 3 convolution filters and max-pooling¹⁹. Two experts [D.Ro., D.Re.] blindly classified the annotations and reviewed the mismatched results to reach a consensus on the group assignment. This collection currently represents to the best of our knowledge the largest dataset in terms of the number of annotated mitotic figures (https://doi.org/10.5281/zenodo.18337703). Therefore, it provides researchers with new opportunities for the development and refinement of data-driven algorithms for identifying mitotic figures on slides and microtiter plates as shown recently for AI-based pattern interpretation of ANA image analysis²⁰. This dataset could also be applied to the segmentation of cells in fluorescence tissue staining when the detection of cell nuclei is of interest.

Methods

Selection and preparation of samples

Immunofluorescence images of HEp-2 cell preparations on glass slides stained with DAPI were obtained from the diagnostic archive of the Association for the Advancement of Immune Diagnostics (Dresden, Germany). The samples originated from 10 years of measurement of consecutive HEp-2 preparations as part of routine diagnostics for the detection of ANA.

Images of wells on the slide were taken with an inverted fluorescence microscope (akironNEO Medipan, Dahlewitz, Germany) in one focal plane per image at a magnification of 200 (Data acquisition parameters: camera model: Basler AC1920-40; sensor: Sony IMX249 CMOS, sensor size: 1/1,2”; image format: uint8, 1936 × 1216, no binning, saved as single channel webp; pixel size: 0.46510 µm/pixel; objective: Plan N 20x/0.4; exposure time: Basler, autoadjusted to staining intensity, range of 30–100 ms; light source: LED 400 nm; filter: triple-band, excitation: 382/473/630, dichroic mirror: 430/529, emission: 405/495/655)^8,21. Cell segmentation was performed using Otsu’s thresholding as implemented in the akironNEO® software, a widely used and well-established method for separating foreground objects from background in bioimage informatics workflows²².

Initial manually annotated dataset by experts - expert dataset (ExpD)

Primary annotations were performed by two experts trained in the field of immunology [R.Hi., D.Ro.]. The experts classified the stained images based on previously reported DAPI image data of mitotic figures into the following groups (Fig. 1)²³:

1. Interphase: Rounded nucleus of mostly homogeneous texture. In individual cases, nucleoli can be seen in the nucleus. Interphase nuclei are characterized by less condensed chromatin, resulting in low intensity staining and larger nuclei compared to the cell division phases.

2. Prophase: Clearly condensed chromatin with an irregular texture, which leads to an increased intensity of the nucleus with smaller nucleus size compared to interphase cells. Prophase nuclei have a rounded shape with distinct corners and edges around the rim.

3. Metaphase: Narrow, linear shape with very intense staining due to highly condensed chromatin in the metaphase plate. Homogeneous to slightly granular texture with some slight corners and edges in the rim.

4. Anaphase: Rounded to rectangular, mostly symmetrical shape with high intensity and recognizable separation zone in the middle area. Protein filaments are partially visible, which pull the chromatin towards the poles.

5. Telophase: Mostly symmetrical, parallel, already separated daughter cells with similar shapes and highly condensed chromatin with intense staining. The stained chromatin is usually smaller than in metaphase.

6. Artefact: Unassignable objects due to detached cells, strong overlap or occlusion of cells during cell culture. Other fluorescent objects that cannot be assigned to any other group.

7. Triple: Cells with discernible division defects, especially in metaphase and telophase. In metaphase, more star-shaped with 3 arms than linearly elongated. In telophases, three instead of two identical daughter cells of similar size. Objects caused by division errors and polyploidy of the tumor cell line.

Fig. 1.

Image examples for the 5 cell cycle phases annotated in the data set (https://doi.org/10.5281/zenodo.18337703) and grouped accordingly¹⁸. Additionally, ambiguously centered cells and overlapping or occluded structures are annotated as artefacts and assigned to one group. Furthermore, triple structures, which occur as a result of cell division errors and are a common feature of cancer cells, were assigned to the Triple group. The last column shows schematic images of the 5 cell cycle phases.

The group of artefacts plays a special role in this context. This group was originally used for images that do not exhibit mitotic figures but cannot be clearly assigned to interphase cells either.

Cells with disagreement were re-assigned to both experts for a common consensus. Both experts performed blinded annotations not knowing cell-class decisions of the other expert (Fig. 2). The annotation algorithm provides a mode for annotating cells that allows the expert to evaluate the annotation of cells with unknown class. After selecting the respective classes, the next random annotation(s) was presented. The resulting confusion matrix, along with the corresponding global metrics, is given in Suppl. Table 1.

Fig. 2.

Creation of the initial manually labelled expert-dataset variant, which is the base for the final CNN-expert data set. Every 20x microscope image was screened automatically for brightest objects to enrich cells of interest. The first and second experts were able to view segmented cell images showing the segmented object in the center and the surrounding objects.

It is well known that the concordance of different experts in assessing mitotic figures is not perfect. In all cases where both experts could not agree on the same classification encompassing several doubtful candidates, both experts re-evaluated them to find agreement on the group. The re-evaluation resulted in the manually labeled dataset (ExpD) variant encompassing 2,500 interphase cell images and 1,250 images each for the remaining groups (Table 1). To lower the risk of missing labels, which tend to happen with manual annotations of large datasets, we used an algorithmic-assisted pipeline to grow our annotated dataset by a factor of ten.

Table 1.

Number of images assigned to the 7 classification groups and corresponding file numbers and data sizes in the expert (ExpD) and enlarged datasets (CNN-D).

Classification groups	ExpD	CNN-D
Interphase	2,500	28,639
Prophase	1,250	6,604
Metaphase	1,250	31,582
Anaphase	1,250	3,245
Telophase	1,250	20,941
Artefact	1,250	6,468
Triple	1,250	2,521
Total file number	10,000	100,000
Size	135 MB	2,960 MB

Enlarged dataset for mitotic figures (CNN-D)

To improve the quality of our dataset, we used a deep learning image classifier based on CNNs trained on the manually labeled ExpD dataset. Based on these data, our primary goal was to improve the separation of similar mitotic figures belonging to sequential mitotic phases, e.g., the transition from prophase to metaphase and from anaphase to telophase.

For cell classification, we used a standard CNN architecture based on VGG16 as a backbone. We trained this network for 25 cycles using the ExpD dataset, utilizing 256 × 256-pixel image slices around the annotated cells in the dataset for each cycle. Predicted cell classifications by the network were reviewed again by both experts to account for potentially misclassified images. Afterwards, the trained VGG16 network was used to pre-classify new images of the same size (n = 100,000).

The goal of this approach was to eliminate the bias and improve the entire dataset using a deep learning image classifier based on CNNs. In this context, a deep network suggests additional potential candidates for mitotic figures, which would need to be evaluated by human experts and assigned to the various classification groups in our dataset (Fig. 3). Eventually, the human experts reviewed the pre-classified images and jointly assigned them to the appropriate groups (Table 1). The resulting confusion matrix, along with the corresponding global metrics, is presented in Suppl. Table 2.

Fig. 3.

Neural network-aided classification of additional images for the enlarged dataset CNN-D (CNN + expert labelled sets).

Data Records

We provide the original images in WebP format on Zenodo (https://doi.org/10.5281/zenodo.18337703)^18,24. All images have been fully anonymized. WebP was used because it maintains a good human-perceivable visual quality and superiority over PNG in lossless image compression²⁵. We use lossless WebP to preserve exact pixel values while benefiting from compact file sizes, which reduces disk usage and accelerates data loading for machine learning training and evaluation^26,27. Each filename contains the group name assigned by experts followed by a unique image number per group (e.g., Interphase_00010) or “DAPI” followed by a unique image number for overall microscopic images.

Technical Validation

Our technical validation of the dataset was performed at two levels: First, we evaluated the quality of the assigned labels by performing a classification experiment of mitotic figures versus other cells. Second, we performed a class recognition task on the complete test set. Both are informative and address different questions: While the first test provides information on how well class separation is possible and, thus, indirectly assesses the quality of label classes, the second test assesses the coverage of mitotic figures in the CNN-D. Accuracy was used as measuring criterion, defined as ratio of the number of correctly predicted images to the total number of evaluated images (Suppl. Tables 1-2).

Classification of preselected cells

For this validation task, 256 × 256 pixel patches containing individual cells of all classes in the center of the image patch except ambiguous cells (mitotic figure) were segmented from the CNN-initial variant of the dataset. We used a state-of-the-art CNN classification network based on a non-pretrained VGG16. The network was trained for a cycle of 25 epochs with a maximum learning rate of 10⁻⁶ and the Lion optimizer. Using this approach, we achieved an accuracy (proportion of correctly predicted images to the total number of images evaluated) of 92.0% with the test set. The main confusion relates to adjacent mitotic figures, while all other cell types were well separated by the classifier. This result is also consistent with the high intra- and inter-rater variance on this task by human experts.

Supplementary information

Supplemental Material ^{(496.2KB, pdf)}

Author contributions

R.Hi. and D.Ro. wrote the manuscript and performed the main research and analysis tasks of this work. R.Hi. performed the data analysis, network training, and method development. K.Co. and D.Re. provided the basic image data and reviewed the cells as immunological experts. P.Sc. contributed to method development and reviewed the manuscript. S.Rö. checked and supported technical aspects, contributed to method development and reviewed the manuscript.

Funding

Open Access funding enabled and organized by Projekt DEAL.

Code availability

The dataset is available at https://zenodo.org/records/18337703¹⁸.

Declarations

Competing interests

D.Ro. is an employee and shareholder of Medipan GmbH and GA Generic Assays GmbH. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41597-026-08216-w.