A Data-Centric Approach to HEK Cell Microscopic Image Segmentation using Multi-Scaling U-Net

Abstract

Accurate and reproducible cell culture monitoring is important in biomedical research and regenerative medicine, yet manual assessment of cell confluency and morphology remains subjective and prone to inter-observer variability. Although deep learning approaches have been widely applied to cell segmentation, their systematic application to Human Embryonic Kidney (HEK) cells using data-centric methodologies remains underexplored. This study addresses this gap by implementing a Multi-Scaling U-Net (MSUNet) architecture combined with a data-centric workflow that emphasizes improving data quality through Contrast Limited Adaptive Histogram Equalization (CLAHE) preprocessing, Total Variation Denoising (TVD), and iterative expert-guided annotation refinement. The scope of analysis was limited to HEK293T cell images captured at 10x magnification using phase-contrast microscopy. The optimized model achieved an Intersection over Union (IoU) of 0.8980 after applying the data-centric approach, representing a 15.1% relative improvement over the baseline model trained without preprocessing or annotation refinement. These findings provide empirical evidence that systematic data quality improvement constitutes a key contributing factor to segmentation performance, offering a reproducible methodology for automated cell confluency measurement in resource-constrained laboratory settings.