Deploying patch-based segmentation pipeline for fibroblast cell images at varying magnifications

• Main Authors: Malik, Hafizi, Idris, Ahmad Syahrin, Toha @ Tohara, Siti Fauziah, Idris, Izyan Mohd, Daud, Muhammad Fauzi, Tokhi, Mohammad Osman
• Format: Article
• Language: English; English
• Published: IEEE 2023
• Subjects: T10.5 Communication of technical information
• Online Access: http://irep.iium.edu.my/111701/7/111701_Deploying%20patch-based%20segmentation%20pipeline.pdf; http://irep.iium.edu.my/111701/8/111701_Deploying%20patch-based%20segmentation%20pipeline_Scopus.pdf; http://irep.iium.edu.my/111701/; https://ieeexplore.ieee.org/document/10239394
• Institution: Universiti Islam Antarabangsa Malaysia

Cell culture monitoring necessitates thorough attention for the continuous characterization of cultivated cells. Machine 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. Thus, the main objective of the research is to propose a segmentation pipeline specifically for fibroblast cell images on phase contrast microscopy at different magnifications and to achieve reliable predictions during deployment. The research employs patch-based segmentation for predictions, with U-Net as the baseline architecture. The proposed segmentation pipeline demonstrated significant performance for the UNet-based network, achieving an IoU score above 0.7 for multiple magnifications, and provided predictions for cell confluency value with less than 3% error. The study also found that the proposed model could segment the fibroblast cells in under 10 seconds with the help of OpenVINO and Intel Compute Stick 2 on Raspberry Pi, with its optimal precision limited to approximately 80% cell confluency which is sufficient for real-world deployment as the cell culture is typically ready for passaging at the threshold.