Oct 2026· Microscopy research and technique (Print)· 0 citations· 24 references
Medicine
TL;DR
EPMSA-Net, a lightweight instance segmentation network that integrates edge priors and multi-scale attention for B-cell image analysis achieves competitive segmentation performance with a compact model size on the current dataset.
Abstract
Precise single-cell analysis is important for life sciences and precision medicine. In antibody drug discovery and immunotherapy, automated segmentation and quantitative analysis of suspension-cultured B cells in label-free bright-field microscopic images can support image-based characterization of cell populations under the recorded culture conditions. However, accurate segmentation remains challenging because of weak boundary gradients, heterogeneous cell sizes, background interference, and high-density clustering. Existing deep learning models may also face a trade-off between segmentation performance and model compactness, which can limit their use in resource-constrained automated microscopy systems. To address these challenges, we propose EPMSA-Net, a lightweight instance segmentation network that integrates edge priors and multi-scale attention for B-cell image analysis. The proposed framework incorporates two complementary structures: the Edge-Aware Feature Preservation (EAFP) module supplements weak boundary-related and local spatial cues during early feature extraction, whereas the Multi-Scale Contextual Attention (MSCA) module subsequently models multi-scale contextual information and recalibrates downstream feature responses. Evaluation on the current suspension-cultured B-cell dataset yielded an F1-score of 94.44% ± 0.53% and an Aggregated Jaccard Index (AJI) of 89.87% ± 0.78% over five independent training runs, with a parameter count of 3.276 M. Density-stratified evaluation showed relatively small F1-score variation across the predefined low-, medium-, and high-density subsets, with a mean CVF1 of 0.84% ± 0.13%. The resulting instance masks were further used to calculate image-level cell count and segmented area ratio (SAR). These results indicate that EPMSA-Net achieves competitive segmentation performance with a compact model size on the current dataset. Further validation across independent culture batches, imaging systems, and hardware platforms is required to evaluate its broader applicability and deployment performance.
The rapid advancement of cellular immunotherapy has generated an urgent demand for the high-throughput and high-precision quantitative analysis of T-lymphocyte functions, the realization of which hinges on the accurate segmentation of individual cell instances within complex microscopy images. However, existing methods...
Nucleus instance segmentation and counting are essential for quantitative microscopy, enabling reliable analysis of cell populations, tissue organization, and disease-related changes. However, accurately separating touching nuclei remains difficult, and many existing approaches rely on complex models or sensitive post-...
Mohammad Sarhangzadeh, E. Esmaeilzadeh, D. Oner· 0 citations
The results show that PM-CResUnet improves instance separation and counting performance while remaining lightweight, and that multi-scale features, boundary constraints, and data diversity act complementarily, with the full model achieving the best overall performance in both instance segmentation and counting.
Zhuang Wang, Pei-Wen Yan· Journal of Computer Science...· 0 citations
Cell segmentation in time-lapse microscopy images is crucial for cell behaviour analysis, and it is required for biological research and medical applications. However, it is a challenging task because of the low contrast, halo artifact, and frequent cell overlap. Manual cell segmentation is not only time-consuming but...
Yusuf Ata, Muaz Mearri, B. Mayalı et al.· 2026 11th International Conf...· 0 citations
A multimodal pseudo-ground truth training technique to enable supervised feature learning on H&E images for whole-cell segmentation and adapted into a ST pipeline to create an integrated end-to-end machine learning solution, without additional optimisation on ST data.
A.-A. Zhang, Ulrich D. Kadolsky, A. Saxena· 0 citations
Predicting gene expression from H&E-stained histology images offers a scalable alternative to costly spatial transcriptomics, yet most existing methods operate at the spot level, where signals from multiple cells are aggregated and critical cellular heterogeneity is obscured. Extending this paradigm to single-cell reso...