3D Bioprinting of Cultivated Meat Followed by the Development of a Fine-Tuned YOLO Model for the Detection and Counting of Lipoblasts, Fibroblasts, and Myogenic Cells.

Abstract

Background: Cultured meat holds significant potential as a pivotal solution for producing safe, sustainable, and high-quality protein to meet the growing demands of the global population. However, scaling this technology requires innovative bioengineering approaches integrated with software methods to assess the growth of cell cultures. This study aims to develop a technology for 3D printing a hybrid meat product and subsequently design a finely tuned You Only Look Once (YOLO) model for detecting and counting lipoblasts, fibroblasts, and myogenic cells.

Methods: Cultures of multipotent mesenchymal stem cells (MMSCs) and fibroblasts were obtained from the domestic rabbit Oryctolagus cuniculus domesticus. Standard protocols were employed to induce adipogenic and myogenic differentiation from MMSCs. Fibroblasts were isolated from skin biopsy samples. The 3D printing process utilized bioinks. The engineering approach involved the development of a unique print head integrated into a 3D printer. Confocal and transmission electron microscopy of the cells within the construct was performed. A dataset of digital images of lipoblasts, myogenic cells, and fibroblasts was created. Four models based on the YOLOv8-seg architecture were trained on annotated images, implemented in the Telegram bot.

Results: Stable cultures of lipoblasts, myogenic cells, and fibroblasts were obtained. 3D-printed tissue constructs composed of rabbit cells, sodium alginate, and sunflower protein were successfully fabricated. A unique print head for a 3D printer was assembled. Confocal microscopy confirmed cell viability within the tissue construct. Ultrastructural analysis revealed dense intercellular contacts and high metabolic activity. The resulting product replicated the organoleptic and structural properties of natural meat. In the IT segment, the single-class model trained on lipoblasts achieved metrics of recall 85%, precision 77%, and mean Average Precision at IoU threshold 0.50 (mAP50) 79%, which improved in the multiclass model to recall 92%, precision 92%, and mAP50 81%. The IT solution was implemented in a Telegram bot capable of detecting and counting different cell types.

Conclusions: A 3D tissue construct was achieved. Detailed microscopic analysis demonstrated cell viability and high metabolic activity within the polymerized alginate hydrogel. The engineered tissue product presents a potential alternative to natural meat. Additionally, the trained neural network models, implemented in a Telegram bot, proved effective in monitoring culture growth and identifying cell types in digital images across three cell cultures. As a result, we developed four YOLOv8 models and demonstrated that the multiclass model outperforms the single-class model. However, all models exhibited reduced accuracy in high-density cultures, where overlapping cells led to undercounting.