Automated confluency of fibroblast skin cell detection using AI-based image segmentation

Abstract

Fibroblast cell culture monitoring necessitates thorough attention for the continuous characterization of cultivated cells. Deep learning has recently emerged to engage in a process, such as a microscopy segmentation task; however, the trained data may not be comprehensive for other datasets. Most algorithms do not encompass a wide range of data attributes and require distinct system workflows. To address this, the research proposed a U-Net based pipeline specifically for detecting and segmenting fibroblast cell confluency at varying magnifications and distributions. U-Net is a widely recognized deep learning architecture commonly employed for biomedical segmentation tasks. Patch-based segmentation was employed for predictions, with three U-Net based networks proposed using inception, dilation, and residual mechanisms. Additionally, a smooth blending technique was introduced to address edge effects arising from patch-based segmentation. Model B, with an inception mechanism of various kernel sizes, demonstrated the highest overall IoU score of 0.707, with a mean relative performance of 6.5% across all datasets. It also outperformed human observation by 15% more accuracy through validation analysis. However, when disregarding deployment considerations, Model C, featuring a dilation mechanism, emerges as a viable alternative, offering optimal performance and moderate complexity compared to Model B during training sessions. Another main objective of this study is to fully automate cell monitoring, traditionally involving high human interaction. Thus, a complete design of automated cell culture monitoring was also designed for a specified working space incorporating AI-based cell screening, IoT-based monitoring and communication system through Cloud Firebase, and sample handling automation using robotic arm. OpenVINO and NCS2 were used to optimize the segmentation on Raspberry Pi where a cell image can be segmented in under 9 seconds by Model B. Overall, the proposed design for automated cell culture monitoring demonstrates its efficiency, with the complete cycle of its workflow taking under 1 minute and an average power rate of 12.2W, while requiring just approximately 200MB of downloaded data for a full day of operation in a real-time application. However, the high precision of the proposed model is currently limited to around 80% confluency and the implementation of the automation system has yet to be adjusted to accommodate varying lab structures. The primary contribution of this study revolves around providing guidance in refining existing models to adapt diverse magnification levels of microscopy images under specific conditions. Furthermore, the integration of Raspberry Pi and OpenVINO in biomedical tasks, a relatively unexplored area, adds a distinctive aspect of this research. Consequently, this study introduces a modular, comprehensive, and optimized cell imaging system with the ability to integrate various external systems, which is off the shelf, enabling cost-effective automated cell culture monitoring.