We used tissues from a donor animal. An adult male rabbit Oryctolagus cuniculus domesticus of the Soviet chinchilla breed served as a donor animal. For the isolation of MMSCs, fragments of adipose tissue weighing 2 grams were obtained from the large omentum by laparoscopic biopsy; for the isolation of fibroblasts, biopsy specimens of the abdominal skin were obtained. All cell lines were validated by short tandem repeat (STR) profiling and tested negative for mycoplasma.

Rabbit was chosen as a donor in this study because this animal is equally successful in meat farming and as a model animal in scientific research. The use of biopsy allowed the life of the donor animal to be preserved, which is one of the most important principles of the cultured meat concept.

All surgical operations are performed under anesthesia. We used anesthesia for the rabbit following this protocol: intramuscular injection of a mixture of Xyla (0.2 mL/kg, 2% xylazine hydrochloride solution; manufactured by Interchemie Werken “de Adelaar” BV, Venray, Netherlands) and Zoletil (15 mg/kg, a mixture of tiletamine hydrochloride and zolazepam hydrochloride, Virbac, Carros, France). The anesthesia was verified by the absence of response to pain stimuli (paw pinch) and suppression of the corneal reflex.

The animals were kept in standard cages with free access to food and water under standard conditions: 12 light/12 dark cycle, temperature 22–25 °C, air exchange rate of 18 changes per hour. International, national and/or institutional recommendations for the care and use of animals were followed. The studies were carried out in accordance with the requirements of the Council Directive of the European Communities 86/609/EEC on the use of animals for experimental research (November 24, 1986), the “Rules of Laboratory Practice in the Russian Federation” (Order No. 708n of August 23, 2010) and the organization of procedures for working with laboratory animals (GOST 33215–2014) and Protocol No. 2, approved by the Bioethics Commission of the Don State Technical University on February 17, 2020.

MMSCs were isolated from a fragment of adipose tissue of the greater omentum. All stages of cell isolation were performed under sterile conditions in a biological safety box. The cell fraction was obtained according to the protocol proposed by Bunnell et al. [45], which, in brief, included incubation of the biopsy specimen in 0.2% collagenase solution from crab hepatopancreas (Biolot LLC, St. Petersburg, Russia) in DPBS (Biolot LLC, St. Petersburg, Russia) for 60 minutes at 37 °C with constant shaking, filtration of the obtained suspension through a cell sieve and cell sedimentation by centrifugation. The obtained cell mass was introduced into culture vials with DMEM/F12 medium (Biolot LLC, Russia) supplemented with 10% fetal bovine serum (Biolot LLC, Russia) and 1 ng/mL recombinant basic human fibroblast growth factor (FGF2, Sigma-Aldrich, St. Louis, MO, USA). After 48 hours, the medium was changed to remove unattached cells. The resulting cell culture was maintained for five passages. Antibiotics and antimycotics were not used during cell culture hereafter. Cell disaggregation during passages was performed using 0.25% Trypsin-Versen 1:1 solution (Biolot LLC, Russia).

For induction of adipogenic differentiation we used DMEM medium with high glucose content and stable L-glutamine (Biolot Ltd., Russia), 10% fetal bovine serum (Biolot Ltd, Russia) 0.1 µmol dexamethasone (D4902, Sigma-Aldrich, USA), 0.45 µmol isobutylmethylxanthine (I5879 Sigma-Aldrich, USA), 170 nmol insulin (I5500 Sigma-Aldrich, USA), 0.2 µmol indomethacin (I7378, Sigma-Aldrich, USA). On the 10th day of culturing, the medium was replaced with complete DMEM medium, which was used in the following steps. As a differentiation control, the MMSC cell culture was cultured in complete DMEM medium without inducers. Lipid droplets detected by microscopy in cell cytoplasm were accepted as a specific marker of differentiation [46].

DMEM medium with high glucose content and stable L-glutamine (Biolot LLC, Russia) was used for induction of myogenic differentiation, 10% fetal bovine serum (Biolot LLC, Russia) 10 mM 5-azacytidine (Sisco Research Laboratories, Mumbai, Maharashtra, India), 50 µmol hydrocortisone (H6909 Sigma-Aldrich, USA) [47]. On the 15th day of culturing, the medium was replaced with complete DMEM medium, which was used in the following steps. As a differentiation control, the MMSC cell culture was cultured in complete DMEM medium without inducers. Cell fusion with the formation of multinucleated myotubes detected by microscopy was taken as a specific marker of differentiation [25].

Fibroblasts were isolated from skin biopsy specimens according to the protocol proposed by Abade dos Santos et al. [48]. All cell isolation steps were performed under sterile conditions in a biosafety box. Tissue fragments were washed five times in DMEM medium with antibiotic and antimycotic solution (200 U/mL penicillin, 200 µg/mL streptomycin, 0.5 µg/mL amphotericin B and 50 µg/mL gentamicin) under vigorous shaking with medium change every 5 minutes to prevent microbial contamination. Tissue fragments were sterilely cut with scissors and placed in trypsin dissociation solution preheated to 37 °C and incubated at 37 °C for 10 minutes, after which the solution was centrifuged at 150 g for 10 minutes. The settled cells were suspended in DMEM medium with 10% fetal bovine serum and antibiotic and antimycotic solution: 100 U/mL penicillin, 100 µg/mL streptomycin, 0.5 µg/mL amphotericin B (Gibco, New York, NY, USA) and 25 µg/mL gentamicin (Biolot LLC, Russia), transferred to culture vials and incubated at 37 °C with 5% CO₂ for 4 hours, after which the vials were washed and a new portion of medium was added. After 2 hours, the washing procedure was repeated and the medium was changed. Subculturing was carried out after reaching 90% cell layer density, and antibiotics and antimycotics were not used in the following passages.

The cell isolation and differentiation protocols used allowed us to obtain stable cell cultures that were subcultured for 25 passages for fibroblasts and lipoblasts and 12 passages for myogenic cells.

In the biotechnology of creating bioinks for 3D printing, a mixture based on sodium alginate and sunflower protein concentrate was used. The dry sunflower protein concentrate (SPC), containing 840 g/kg of protein (Kjeldahl method), employed in this study was supplied by LLC “MEZ YUG RUSI” (Rostov-on-Don, Russia).

All components were sterilized under a UV lamp in a specialized biosafety cabinet for 40 minutes prior to use to prevent potential microbial contamination. Sodium alginate (Sigma-Aldrich, St. Louis, MO, USA) at a concentration of 30 mg/mL was dissolved in Dulbecco’s Phosphate-Buffered Saline (DPBS) on a rotary shaker at room temperature for 1 hour to ensure complete dissolution and homogeneity of the solution. The next step involved incorporating the sunflower protein concentrate, maintaining a protein-to-alginate mass ratio of 3:1 based on the dry weight of the substances. The resulting mixture was thoroughly mixed until a uniform paste-like consistency was achieved. To ensure even distribution of the components, the mixture was allowed to rest at room temperature for 2 hours. After preparing the paste-like hydrogel, 3 mL of the mixture was transferred into a sterile 5-mL syringe, which was centrifuged at 3000 rpm for 3 minutes to remove air bubbles, a crucial step for ensuring the quality of the subsequent bioprinting process.

Immediately before printing, the hydrogel was enriched with cells. Fibroblasts, adipocytes, and myosatellite cells were mixed in a 1:1:1 ratio and pelleted via centrifugation. The pelleted cells were resuspended in the hydrogel at a concentration of 2 $\times$ 10⁶ cells per gram of hydrogel, using a mixing technique involving two syringes connected by an adapter to achieve uniform cell distribution within the matrix. The prepared bioink was then loaded into an injector mounted on the print head of a modified Anet A8 3D printer (Shenzhen Anet Technology Co., Ltd., Shenzhen, China). After completing the setup, the bioprinting process commenced.

For 3D bioprinting, we designed and constructed a unique print head equipped with a hydraulic drive in the form of a peristaltic pump and a stepper motor. This drive system synchronized the extrusion speed of the material with the motion of the 3D printer’s actuators and allowed for reverse extrusion at the end of each layer, improving print quality. The Slic3r software, included in the Repitier Host 3D printer control package (Hot-world GmbH & Co, Willich, North Rhine-Westphalia, Germany), was used to prepare the model with the following parameters: object dimensions – 30 $\times$ 40 $\times$ 3 mm, layer height – 0.2 mm, number of perimeters – 0, number of solid layers at the top and bottom – 0, infill density – 8% with Rectilinear structure, infill angle – 90°, infill printing speed – 10 mm/s. A nozzle with a diameter of 0.2 mm (25G) and a filament diameter of 1.2 mm were used, with an extruder feed rate coefficient of 1.3.

Printing was performed in a laminar flow environment within a biosafety cabinet at room temperature. The material was deposited in 60-mm Petri dishes using air as the medium for structure formation. After printing, the constructs were transferred to a CaCl₂ solution (100 mg/mL) and incubated for 10 minutes to crosslink the alginate component. The crosslinked structures were then placed in Petri dishes with DMEM culture medium and incubated in a CO₂ incubator at 37 °C for 72 hours.

Sytox Green Stain (ThermoFisher Scientific, Waltham, MA, USA), a highly sensitive fluorescent indicator for cell nuclei, was used to analyze tissue constructs by confocal laser scanning microscopy. The working solution of the dye was prepared in a 1:1000 dilution based on sterile phosphate buffer (PBS). Fragments of tissue constructs were immersed in the prepared solution and incubated for 20–30 minutes at room temperature under dark conditions, which ensured uniform staining of cell nuclei containing the damaged membrane.

After incubation, the samples were thoroughly washed in several changes of PBS to remove excess dye and minimize background signal. Then the constructs were placed in a medium to prevent signal photobleaching (Abberior, Göttingen, Lower Saxony, Germany), which protects fluorescent molecules from photodegradation during the study. This is especially important for long-term visualization of highly detailed structures. The samples were covered with a cover glass to prevent drying and improve optical properties during microscopy.

The samples were examined using an inverted confocal laser scanning microscope Abberior Facility Line (Abberior Instruments GmbH, Göttingen, Lower Saxony, Germany), which allows for detailed analysis of cellular and tissue structures with high spatial resolution. To visualize the 3D object, the object was scanned along the Z axis with a step of 200 nm with a pixel size of 40 nm. The 3D model was constructed using the ImageJ program (version 1.54j, National Institutes of Health, Bethesda, MD, USA).

Tissue construct fragments were fixed in 2.5% glutaraldehyde (Aurion, Eugene, OR, USA) to preserve their structure. After that, the samples were washed in PBS to remove excess fixative and subjected to additional post-fixation in 1% osmium tetroxide (OsO₄) in phosphate buffer for 1.5 hours.

To prepare for embedding in epoxy resin, the samples underwent a dehydration stage: they were successively treated with alcohols of increasing concentration, bringing them to absolute ethanol. Then, three stages of propylene oxide treatment were used for intermediate infiltration, after which the samples were embedded in Epon-812-based epoxy resin to create a stable matrix. Semi-thin and ultra-thin sections were prepared using an EM UC 7 ultramicrotome (Leica, Wetzlar, Hesse, Germany) and an Ultra 45° diamond knife (Diatome, Biel/Bienne, Canton of Bern, Switzerland).

Ultra-thin sections were contrasted with uranyl acetate and lead citrate solutions to enhance the detail of intracellular structures. They were then analyzed using a JEM-1011 transmission electron microscope (Jeol, Akishima, Tokyo, Japan) at an accelerating voltage of 80 kV, which allowed high-resolution images to be obtained.

Semi-thin and ultra-thin sections were prepared using an EM UC7 ultramicrotome (Leica, Germany) and an Ultra 45° diamond knife (Diatome, Switzerland), ensuring high-precision sectioning with minimal artifacts. Semi-thin sections, 0.5–1 µm thick, were stained with a 1% aqueous solution of methylene blue to enhance contrast and visualize the structural elements of the tissues. After staining, the sections were carefully rinsed with distilled water to remove excess dye and air-dried.

The examination of semi-thin sections was performed using a modern light-optical microscope, Leica DM6000 B (Leica Microsystems, Germany), equipped with a high-precision optical system and a digital camera, enabling high-resolution image documentation. The use of this equipment provided a detailed study of morphological changes in tissues, including the analysis of cellular and extracellular structures.

In this study, we analyzed cells that typically grow as a monolayer within culture flasks. One of the primary challenges of visualizing such cell cultures is their low contrast and reduced natural coloration due to the absence of biological fluids. Unlike fluorescent microscopy, where cell boundaries and intracellular structures are well-defined through the use of various stains, images of cultivated meat cells pose significant difficulties for human operators and software-based methods for analyzing digital images of biological structures, including neural network algorithms.

Cell counting was performed directly in the culture flask, which allowed for maintaining sterility and avoiding unnecessary manipulations that could negatively affect cell viability. This approach minimizes the risk of contamination, saves time, and reduces cell material loss, ensuring more accurate results. Digital images of monolayer cells in culture flasks were captured using an inverted optical microscope (Leica DM IRB) equipped with a digital CCD camera (5 MP) via the ToupView software (version 4.11, Hangzhou, Zhejiang, China) at $\times$100 magnification in phase contrast mode (Supplementary Fig. 1)

The resulting dataset included three types of cell cultures: myogenic cells, fibroblasts, and lipoblasts. It represented a heterogeneous collection of digital images characterized by variability in image quality, morphological features, and cell density. The visualized cells exhibited low contrast, indistinct boundaries, and complex shapes, making their identification and annotation highly challenging.

For manual object annotation in the microphotographs, the Make Sense service (online image annotation service, developed by Piotr Skalski, Poland) was utilized in polygonal annotation mode, and annotation files were saved in comma-separated values (CSV) format (Fig. 1a–c). For additional manual counting of images without annotation, the operator employed the ImageJ application [49].

Fig. 1.

Cell markings on micrographs in the Make Sense service and workflow of automated cell detection using deep learning. (a) Lipoblasts. (b) Myogenic cells. (c) Fibroblasts. (d) Workflow of Automated Cell Detection Using Deep Learning. Scale bar 200 µm. The process begins with an original microscopic image, followed by cell annotation to highlight the structures of interest. The annotated image undergoes preprocessing, dividing it into smaller tiles for improved training efficiency. These preprocessed images are then used for model training, enabling the detection model to recognize cell structures. The final result displays the detected cells, marked in red, with a total cell count indicated in inference route.

The created dataset represents a complex and heterogeneous resource for cell analysis, necessitating the development or adaptation of specialized algorithms for automated processing.

The dataset consisted of 1022 bitmap (BMP) images with a size of 2592 pixels by 1944 pixels, which included 56 images containing segmented annotations for all three cell cultures with different cell densities (Supplementary Fig. 2). The total number of annotated cells reached 17,509.

Manual cell annotation in photographs was performed using the web service Make Sense. Annotated images were saved in COCO JSON format for subsequent analysis. To enhance model performance, we implemented a preprocessing strategy aimed at improving image contrast and expanding the dataset. Initially, all images were converted to grayscale. Each image was then divided into 25 equal parts using a 5 $\times$ 5 grid with the Roboflow tool (Roboflow, Inc., San Francisco, CA, USA). The annotations were converted from JSON format to YOLOv8-seg format and assigned to the corresponding split images.

Subsequently, we enhanced the image contrast using the cv2.createCLAHE function with parameters set to clipLimit = 15.0 and tileGridSize = (8, 8). To further improve accuracy, we trained specific models for each cell culture separately and a multi-class model that was trained on all three cultures. Each dataset was divided into three groups: 975 images for training, 250 for validation, and 125 for testing. Since YOLO internally applies augmentations such as resizing, rotation, and mosaic during training, no manual augmentations were applied to the input images (Fig. 1d).

In this study, the YOLOv8-seg model was refined, specifically its small variant (YOLOv8S-seg). This procedure was carried out to address the task of segmenting and counting cells in images of cultured cells.

The model was trained for 100 epochs, as the validation metrics indicated an optimal point within this range. During training, the patience parameter was set to 50 to control early stopping, and model weights were periodically saved every 20 epochs. These training parameters were selected based on prior studies, which demonstrated that smaller batch sizes could unexpectedly correlate with more effective model training. While this might seem counterintuitive, this effect is particularly evident in autoencoder neural network structures, which are often applied to analyze complex biological data typical of medical research [50]. Due to the large dataset for the first cell culture and limited computational resources, the decision was made to adapt the model architecture to the YOLOv8S-seg variant and reduce the batch size to 4.

During inference, the confidence threshold was initially set at the standard value of 0.5. However, this resulted in suboptimal object detection. To enhance the model’s sensitivity and increase the number of detectable cells, the threshold was lowered to 0.05. Additionally, the standard value for the maximum number of detections per image was increased from 300 to 7000 to improve the model’s accuracy when working with high cell density.

The study utilized a wide range of computational and software resources to implement the most efficient processes for data analysis and machine learning. The primary development tool was Python version 3.10.12 (version 3.10.12, Python Software Foundation, Beaverton, OR, USA), chosen for its versatility and high-level syntax, making it an ideal fit for the study’s objectives. For computer vision tasks, the OpenCV-Python library version 4.10.0 (version 4.10.0, OpenCV.org, Santa Clara, CA, USA) was used, offering numerous tools for image processing, object detection, and real-time data analysis. NumPy was employed to support complex computations, enabling operations on multidimensional arrays and matrices with high efficiency.

Data exchange between applications was facilitated by the JSON format. Object detection in images was performed using the Ultralytics YOLOv8.2.87 model, based on the YOLO algorithm, which achieved high accuracy and processing speed.

Deep learning models were trained using the PyTorch library version 2.4.1+121 (version 2.4.1+121, Meta AI, Menlo Park, CA, USA), widely used for designing and optimizing neural network architectures. Compatibility between various machine learning frameworks was ensured through the use of the Open Neural Network Exchange (ONNX) format.

A key role in performing computations was played by the NVIDIA GeForce RTX 4090 GPU (NVIDIA Corporation, Santa Clara, CA, USA). This high-performance graphics card was selected for its ability to handle complex deep learning and artificial intelligence tasks, delivering exceptional performance.

Statistical analysis was performed using single-factor analysis of variance (ANOVA) with the posteriori Holm-Bonferroni correction criterion. The normality and uniformity of the variance were assessed using the Shapiro-Wilk and Brown-Forsyth tests, respectively. If the normality or homogeneity of the variance was not confirmed, the nonparametric Friedman criterion was used. All the results of the study were analyzed blindly. The differences were considered significant at p 0.05 and n = 5. The data obtained are expressed as an average value $\pm$ the standard error of the mean. Statistical analysis was performed in the SigmaPlot 12.5 (version 12.5, Systat Software, Inc., San Jose, CA, USA) and JASP 0.19.1 (version 0.19.1, JASP Team, Amsterdam, Netherlands) programs.

In this study, the rabbit was selected as the cell donor. This choice was driven by several significant reasons. Rabbits are commonly used in livestock farming for meat production [51, 52], making them an ideal subject for tissue engineering aimed at creating cultured meat products. Additionally, rabbits are widely utilized as model animals in scientific research [53]. For cell harvesting, a biopsy was employed—a minimally invasive procedure that allows for cell extraction without harming the animal. This method preserved the donor’s life, fully aligning with the ethical and ecological principles of cultured meat production.

Following the biopsy, standardized protocols were implemented for cell isolation and differentiation. These protocols included sequential tissue processing steps, such as enzymatic dissociation, cellular fraction isolation, and cultivation in specialized culture media. These methods enabled the establishment of stable cell cultures with high viability and strong proliferation potential.

The obtained cells were categorized into three main lines: lipoblasts, myogenic cells, and fibroblasts (Fig. 2). Each line possesses unique properties essential for producing various components of cultured meat. Fibroblasts provide structural tissue support [54], adipocytes play a key role in flavor development due to their lipid inclusions [55], and myogenic cells form muscle fibers, the primary component of meat products.

Fig. 2.

Obtained cell lines for growing artificial meat. (a) Lipoblasts. (b) Myogenic cells. (c) Fibroblasts. Scale bar 200 µm. Figure crafted exclusively for this study.

For each cell line, a series of passages was performed to assess their stability and long-term viability. Fibroblasts and lipoblasts were successfully subcultured for 25 passages without significant reductions in proliferative activity or changes in morphological characteristics, demonstrating their resilience to the cultivation process. Myogenic cells also performed well, maintaining stability for 12 passages (Fig. 2).

To create and test the 3D printed tissue constructs, bioinks were used (Fig. 3a) to ensure the stability of the structure and support normal growth of cell populations. One of the key components was sodium alginate, which stabilized the hydrogel through a polymerization mechanism in the presence of calcium ions. This process enhanced the mechanical strength of the construct, creating conditions for long-term cell culture and maintaining their functionality.

Fig. 3.

The process of preparing a 3D construct. (a) Prepared bioink consisting of organic substances and cells. (b) 3D bioprinting using a universal head with a hydraulic drive in the form of a peristaltic pump and a stepper motor. (c) Printed 3D construct enriched with beetroot juice.

An important component of the bioink was sunflower protein concentrate, which underwent preliminary cytotoxicity testing. Studies confirmed that it had no adverse effects on cell viability and did not disrupt the 3D printing process. Furthermore, due to its unique rheological properties, the concentrate provided the texture and density characteristic of natural meat. This component also facilitated the uniform distribution of cells within the hydrogel, which was critical for forming a tissue-like structure. A unique hydraulic printhead equipped with a peristaltic pump and stepper motor ensured flawless 3D printing (Fig. 3b).

To achieve a visual resemblance to meat, raw beetroot juice was added to the bioink (Fig. 3c). This gave the bioink a rich reddish hue that persisted until thermal processing. Beetroot juice also enriched the bioink with bioactive compounds, enhancing its nutritional value.

The finished tissue constructs were subjected to frying to evaluate their organoleptic properties. The cooking process was as follows: the samples were pan-fried on both sides in a preheated open skillet at 200 °C using refined, deodorized vegetable oil. The frying time was 4 minutes, and no spices or flavorings were added to preserve the product’s natural taste.

During thermal processing, the Maillard reaction was observed, wherein the initially red product gradually turned a golden-brown hue. This effect created the appearance characteristic of traditionally fried meat, enhancing both the visual and flavor likeness.

Volunteers invited for tasting evaluated the taste, aroma, and texture of the cooked product. They described the texture as firm and juicy, with a taste and aroma reminiscent of natural meat. The crispy crust of the fried product was particularly noted for reinforcing the perception of authenticity. These findings confirm that the developed bioink and 3D printing technology have the potential to create alternative meat products with high organoleptic qualities.

To assess the state of the cell nuclei and possible signs of damage, light and fluorescent images of tissue construct fragments were obtained and analyzed. This approach allowed us to study the cell morphology in detail, as well as to identify key features of their organization and structure.

In semi-thin sections stained with methylene blue, the construct fragments after 72 hours consisted of spindle-shaped cells with elongated processes and moderately stained cytoplasm distributed around the nucleus (Fig. 4a,b). It was observed that the cells were arranged in round structures resembling rosettes, forming clusters in which they were in direct contact with one another. These cell aggregates were characterized by a dense arrangement of cells, indicating interaction between them and potential cooperation in the formation of tissue structures. The cell nuclei were oval or round in shape with fine dispersed chromatin (Fig. 4c). In some fragments of the construct, areas were observed that appeared as externally empty lumens, representing gaps between cells without cellular material. In other regions, these lumens were filled with cytoplasmic protrusions from neighboring cells, indicating active interaction between the cells and their attempt to form contacts. These protrusions may play a role in maintaining the structural integrity of the construct, as well as in facilitating the exchange of substances and signals between the cells.

Fig. 4.

Light and confocal laser microscopy of a tissue construct sample. (a) Transverse semi-thin section of a fragment stained with methylene blue (scale bar 300 µm; magnification $\times$50). (b) Cellular organization within the construct (scale bar 40 µm; magnification $\times$200). (c) Cellular organization within the construct (scale bar 20 µm; magnification $\times$400). (d) Visualization of cell nuclei within the construct using confocal laser scanning microscopy with Sytox Green staining (scale bar 50 µm; magnification $\times$600). (e) Confocal laser scanning microscopy of a construct fragment stained with Sytox Green. A three-dimensional reconstruction of the cellular organization and nuclear distribution within the construct was performed using the ImageJ software, providing a detailed visualization of the spatial arrangement and structural features of the sample.

When studying the preparations, the formation of nuclear rounded groups was noted, visually resembling spheroids. Such structures are compact rounded conglomerates formed from several dozen cells. Each group of cells was organized in a circle, with nuclei located in the form of a crown around the central part. This can be seen in Fig. 4d,e, where the cell nuclei have a partially ordered organization.

The cell nuclei were oval, slightly elongated, with clear and distinct contours. Uniform distribution of nucleic acids within the nuclei was observed, indicating their stable state. Confocal microscopy did not reveal any signs of fragmentation of cell nuclei, their destruction, or transformation. No structures characteristic of damaged cells, such as fragmentation of nuclear material or abnormal changes in their shape, were detected. All cell nuclei retained normal morphology (Fig. 4c–e), indicating their normal viability.

The data obtained during confocal microscopy emphasize the stability of cellular structures in the tissue construct. Partial organization of cells in the form of compact spheroids with uniform distribution of nucleic acids in the nuclei and a clear structure is also observed, indicating the absence of their damage and a high degree of preservation of the cellular material. These results confirm the success of the tissue cultivation and preparation methods used, and also demonstrate the prospects for further research in this area.

Ultrastructural analysis revealed the presence of cells in the tissue construct characterized by the morphological features of fibroblasts and lipoblasts, exhibiting various shapes, namely spindle-shaped, elongated, and triangular. These cells possessed thin, branched projections predominantly oriented along the axis of the cell body. Additionally, small, irregularly shaped protrusions were observed in the cytoplasm. The cells demonstrated dense packing, closely adhering to one another, which resulted in the formation of regions with direct membrane contact between cell bodies and projections (Fig. 5).

Fig. 5.

Transmission electron microscopy of cells within the construct. (a) Fibroblast; (b) lipoblast; (c) nucleus; (d) nucleolus; (e) rough endoplasmic reticulum; (f) lipid droplets, magnification $\times$10,000, scale bar 2 µm.

Fibroblasts and lipoblasts contained oval nuclei with evenly distributed fine-granular chromatin. In most cells, one or two large nucleoli were clearly visible, indicating the high functional activity of fibroblasts. In some cells, nuclei with a curved shape were also observed, likely associated with invagination and protrusion of the nuclear envelope (Fig. 5).

The cytoplasm of fibroblasts was rich in organelles, indicating their high metabolic activity. Numerous oval and round mitochondria, elements of rough endoplasmic reticulum (RER), Golgi apparatus, lysosomes, centrioles, and multivesicular bodies were discernible. The endoplasmic reticulum (ER) occupied a significant portion of the cytoplasm, consisting of parallel-packed and curved tubules as well as rounded and elongated cisternae with signs of dilation. The membranes of the RER were studded with ribosomes on the cytoplasm-facing side, indicating intense protein synthesis. Free ribosomes, unattached to membranes, were also found in the cytoplasm (Fig. 5).

In contrast to fibroblasts, lipoblasts contained numerous inclusions in the form of round, osmiophilic lipid droplets of various diameters. Particularly notable was the well-developed ER network and highly active Golgi apparatus, suggesting processes of active protein synthesis, including collagen. The ER spaces were organized into stacks of membrane structures and complex tubules. The ER occupied a significant portion of the cytoplasm, consisting of parallel-packed and curved tubules as well as rounded and elongated cisternae with signs of dilation. The membranes of the RER were studded with ribosomes on the cytoplasm-facing side, indicating intense protein synthesis. Free ribosomes, unattached to membranes, were also found in the cytoplasm (Fig. 5).

The dataset included 1022 high-resolution images of cultured cells from three types: lipoblasts, myogenic cells, and fibroblasts. Manual annotations were performed on a subset of these images using Make Sense, providing segmented labels for 17,509 cells across 56 images. These annotations were converted into YOLOv8-seg-compatible formats and used for training, validation, and testing. Additionally, the images were preprocessed to grayscale and enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE).

Preprocessing steps were critical for improving detection accuracy, especially on low-contrast images. Without these steps, both models performed poorly at detecting cells, particularly for fibroblast and myogenic cell cultures, where biological object boundaries were not clearly defined. Segmenting the images into smaller parts significantly enhanced model performance, especially in high-density cultures.

Data augmentation, including built-in methods like rotation and resizing, made the model more robust to morphological heterogeneity of the cells, although additional algorithmic solutions, such as synthetic data generation, were not employed. Implementing more advanced augmentation techniques could potentially improve performance in high-density cultures.

It was shown that the number of lipoblasts (Culture 1) calculated by the operator and neural network models of single and multiclass type significantly differ (p 0.001). At the same time, there are significant differences in accuracy between the models (p 0.05) (Fig. 6). However, the dynamics of the calculation accuracy fluctuated greatly at different time intervals. So after 1, 2, and 3 days, both models showed a high result in the accuracy of cell counting, which practically corresponded to the number of lipoblasts that were identified by the operator. The general dynamics of culture growth was reflected in the calculated single and multiclass models, and increasing inaccuracy was not a critical parameter for assessing cell survival at various time intervals (Fig. 6, Supplementary Table 1). However, on the 5th and 6th days, both models experienced a sharp decline in cell identification, which was probably due to several factors at once, namely, the models were not trained on marked photographs during these time intervals, the accumulation of cells was very dense and the boundaries of many of them were indistinct, which did not allow the programs identify each of them. As a result, several objects were taken into account at once, which led to seriously underestimated estimates of the cell population.

Fig. 6.

Graph of cell counting by the operator and neural network models of mono- and multiclass type for lipocyte culture at different time intervals. Scale bar 200 µm. Single-factor analysis of variance was performed. M $\pm$ SEM. n = 5 for each group in a certain time interval. ***p 0.001 — accuracy of neural network models relative to the control group calculated by the operator; *p 0.05 — accuracy of a single-class model relative to a multiclass model. M, mean; SEM, standard error of the mean.

Our next step was to evaluate the accuracy of the models in calculating the number of myogenic cells (Culture 2) at different time intervals compared to the operator. We have demonstrated that the numbers calculated by the operator and neural network models of mono- and multiclass type differ significantly (p 0.01). At the same time, there are significant differences in accuracy between the models (p 0.01) (Fig. 7). It is worth noting that the multiclass model showed great similarity with the control group. The opposite effect was obtained using a single-class model, which initially could not cope with the task, calculating the negative dynamics. On the 3rd and 4th days, the multiclass model gave severely underestimated results that were inconsistent with the control indicators (Fig. 7, Supplementary Table 2). This is probably due to the fact that the models were not trained on the datasets during these time periods. Also, this depended on the cellular objects, whose boundaries were extremely difficult to identify at late stages of culturing, even manually.

Fig. 7.

Graph of cell counting by the operator and neural network models of mono- and multi-class type for myogenic cell culture at different time intervals. Scale bar 200 µm. M $\pm$ SEM. n = 5 for each group in a certain time interval. **p 0.01 is the accuracy of neural network models relative to the control group calculated by the operator; ^$⁡$p 0.01 is the accuracy of a single-class model relative to a multiclass model.

Next, we evaluated the accuracy of the models in calculating the number of fibroblasts (Culture 3) at different time intervals compared to the operator. We have demonstrated that the numbers calculated by the operator and neural network models of mono- and multiclass type differ significantly (p 0.001). At the same time, there are significant differences in accuracy between the models (p 0.01) (Fig. 8). Both models showed good results in determining the number of cells relative to the control group. It is worth noting that up to the 3rd day, neural networks effectively coped with calculations, reflecting the overall dynamics. Of course, the variability of the values obtained using computer programs was quite high, but in general it did not have critical values for the prognosis of cell culture survival. Then the multimodal model began to show negative dynamics, and the single-class model was able to maintain the accuracy of the results until the 4th day (Fig. 8, Supplementary Table 3).

Fig. 8.

Graph of cell counting by the operator and neural network models of mono- and multi-class type for fibroblast culture at different time intervals. Scale bar 200 µm. M $\pm$ SEM. n = 5 for each group in a certain time interval. ***p 0.001 is the accuracy of neural network models relative to the control group calculated by the operator; **p 0.01 is the accuracy of a single-class model relative to a multiclass model.

Our study demonstrated significant differences in the accuracy of lipoblasts, myogenic cells and fibroblasts counting between operator assessments and neural network models of both single-class and multiclass types. While the single-class and multiclass models showed high cell-counting accuracy during the early culture periods (days 1 to 3), both experienced a marked decline in performance by days 5 and 6, particularly due to dense cell accumulation and indistinct boundaries that hindered proper identification. The multiclass model generally outperformed the single-class model, especially for lipoblasts and myogenic cells, showing greater consistency with the operator’s control results. However, discrepancies were noted, especially for myogenic cells counts, where the multiclass model showed underestimation on days 3 and 4. For fibroblasts, both models performed well up to the 3rd day, after which the multiclass model exhibited a negative trend, while the single-class model maintained accuracy longer. These findings suggest that while neural network models show promise in cell counting, especially during early culture periods, their accuracy can be limited by training data and cell density, particularly in later stages of culturing.

We identified the optimal model, “S,” (Small size) through a rigorous analysis of the confusion matrix, which demonstrated consistently robust performance across all evaluation scenarios. This selection prioritizes an ideal balance between precision and recall, achieving the highest overall efficiency. Additionally, the choice of this model was reinforced by evaluating the mean Average Precision (mAP 0.5) across a range of Intersection over Union (IoU) thresholds exceeding 50%, which further confirmed its effectiveness. The IoU is calculated as:

$$IoU=\frac{AreaofOverlap}{AreaofUnion}$$

The small (S), medium (M), and large (L) model configurations were thoroughly trained; however, with no significant variations observed in the confusion matrices across these sizes (Fig. 9), the “S” model was deemed optimal due to its faster inference speed and computational efficiency. This model also offers the advantage of accessibility on lower-spec servers, making it more versatile for deployment in resource-constrained environments.

Fig. 9.

Confusion matrix for YOLOv8S (a) and YOLOv8L (b) model trained on 100 epochs for cell culture classification. This matrix illustrates the YOLOv8S model’s performance in distinguishing three cell culture types (Culture 1, Culture 2, and Culture 3) and background. The matrix shows the count of true versus predicted labels, with higher values along the diagonal representing correct classifications. Culture 1 achieved high accuracy, though misclassifications are notable between Culture 1 and Culture 3, as well as Culture 3 and background. The model was trained over 100 epochs to enhance detection accuracy across categories. YOLO, You Only Look Once.

The performance of the YOLOv8S-seg model was evaluated in two configurations: single-class and multiclass models. Single-class models were trained and tested on individual cell cultures, while the multiclass model was trained on three different cultures but limited to recognizing specific cell types.

Performance was assessed for each configuration across three cell cultures. The results are presented in Table 1, which shows the recall, precision, F1 score, and mAP50 metrics for the YOLOv8S-seg model, comparing its performance in single-class and multiclass configurations across biological samples (referred to as “crops”).

For Culture 1, the model demonstrated high performance: recall was 0.8500, meaning 85% of target objects were correctly detected. Precision was 0.7733, indicating that 77% of detected objects were accurate. The F1 score, balancing recall and orecision, was 0.8098, while mAP50 reached 0.79135, reflecting strong object localization capabilities. For Culture 2, the single-class model struggled, achieving a recall of 0.5686, although precision remained acceptable at 0.7404, leading to an F1 score of 0.6433. The mAP50 for Culture 2 was 0.51382, indicating lower localization accuracy. For Culture 3, the model showed satisfactory results: recall was 0.7684, precision was 0.6765, F1 score was 0.7195, and mAP50 was 0.68511, reflecting balanced accuracy in detection and localization.

In the multiclass configuration, the model was trained on all three cultures simultaneously but made predictions for individual cell types. For Culture 1, the multiclass model demonstrated excellent results: recall was 0.9187, precision was 0.9169, F1 score was 0.9178, and mAP50 was 0.8170, showing outstanding performance across all metrics. In Culture 2, the multiclass model achieved comparable results to the single-class model, with recall at 0.5621 and a slight improvement in precision to 0.6798. The F1 score was 0.6154, and mAP50 increased to 0.5820, indicating a modest improvement in localization accuracy. For Culture 3, the model achieved recall of 0.7671, precision of 0.6499, F1 score of 0.7036, and mAP50 of 0.7020, showing consistent performance in both configurations for this culture.

Comparison between single-class (Supplementary Fig. 3) and multiclass models shows that the multiclass configuration outperformed the single-class model in Culture 1, demonstrating significant improvement in all metrics. However, for Culture 2 and Culture 3, the benefits of the multiclass approach were less pronounced, with slight improvements in precision, F1 score, and mAP50. These results indicate that, while the multiclass model shows better generalization, particularly for Culture 1, some cultures may benefit more from specialized single-class models.

The mAP50 metric, used as a simple but effective indicator of model performance, measures the overlap between predicted objects and ground truth annotations. It is particularly useful for comparing models in object detection tasks, where precise overlap is crucial. For each image, mAP50 was calculated to quantify detection accuracy.

Fig. 10 presents a detailed comparative analysis of the mAP metric for different model sizes and image preprocessing methods. It effectively illustrates how these factors influence model performance.

Fig. 10.

Comparison of mAP50 performance across model dimensions and preprocessing techniques. The figure presents a detailed comparative analysis of the mean Average Precision (mAP) metric across different model dimensions and image preprocessing techniques, providing insights into how these factors influence the performance of the YOLOv8S-seg model. The figure was generated using Matplotlib (Python 3.12.4), a Python-based data visualization library. The data used in the plot comes from the CSV files that were automatically generated as a result of the YOLO training process, following the appropriate preprocessing steps.

The key criterion for selecting the optimal preprocessing method and model size was the mAP50 metric. Fig. 10 provides a comparison of various preprocessing options, model sizes, and class counts, enabling the selection of the most suitable model with the best metrics.

It was demonstrated that the optimal number of training epochs for low-contrast images is up to 50. Beyond 50 epochs, a plateau in metrics such as recall and precision is observed, making additional training inefficient. This suggests that a performance limit may have been reached, beyond which further training could lead to overfitting or minimal improvements.

The model shows a noticeable decrease in detection accuracy when working with black and white images. This is likely due to the absence of important visual features, such as color gradients and contrasts, which play a critical role in object recognition. The reduction in available visual information in grayscale limits the model’s ability to accurately localize and identify cellular structures.

Data augmentation plays a crucial role in improving model performance. By increasing the heterogeneity and volume of training data, augmentation significantly enhances the model’s ability to generalize results on images with varying contrasts. This process enables the model to better recognize and localize objects across images of different quality, a characteristic typical of biological microphotographs.

The study utilized YOLO deep learning models to perform automated cell segmentation within an image processing pipeline implemented in a Telegram bot interface. The bot supports four model configurations: a general “Multiclass” model and three specialized models trained on specific cell cultures, designated as “Lipoblasts”, “Myogenic Cells”, and “Fibroblasts”. Each model was designed and optimized for processing various cell types with distinct morphological parameters. ONNX was employed for model compatibility, facilitating integration into the Python-based Telegram bot. The bot’s operation is illustrated in Fig. 11.

Fig. 11.

Segmentation of the image of cells of three cultures in different time intervals of the Telegram bot operating on the basis of the trained YOLO model. Scale bar 200 µm.

Upon receiving an image from a user, the bot divides it into 640 $\times$ 640-pixel fragments to ensure compatibility with YOLO models and prevent memory overflow. For images larger than 640 $\times$ 640 pixels, a sliding window method with a 50% overlap parameter is applied, reducing data loss at the edges of each fragment and improving segmentation accuracy. Each fragment is then processed using the selected model. Segmentation is achieved by isolating cells, with bounding boxes and masks delineating individual cells or cell clusters. To reconstruct a continuous segmentation of the original image from the processed fragments, a matrix-based merging algorithm was developed. This algorithm overlays each fragment onto the original image size, averaging values in overlapping regions to account for multiple detections and minimize boundary artifacts.

For every segmented image, the YOLO model provides not only spatial segmentation but also classification of detected cells based on predefined model classes. Upon completion of the segmentation process, the bot sums up the detected cells according to their categories, delivering a summary of the cellular composition in each image. For images processed using the “Multiclass” model, the bot filters detection results to focus on a specific cell type as requested by the user. This allows for a quantitative assessment of particular cell populations in heterogeneous samples. After processing, the bot delivers the annotated image and corresponding cell count directly to the Telegram chat, making the analysis process fast and user-friendly.