Multi-Task Learning Framework for Label-Free Single-Cell Phenotyping
Researchers have developed a comprehensive Deep Learning framework aimed at classifying White Blood Cells and tracking protein expression without the need for labels, using Differential Phase Contrast images. The model features a Hybrid architecture that combines detailed texture analysis from convolutional networks with broader context provided by transformers, utilizing a learnable cross-branch gating module for enhanced analysis. Furthermore, a Large Language Model generates concise, biologically meaningful summaries of the predicted states of the cells. The framework was evaluated using the Berkeley Single Cell Computational Microscopy and Blood Cells Image benchmarks, offering a scalable and non-invasive alternative to traditional fluorescence-based cytometry.
Key facts
- Framework jointly performs WBC classification and protein-expression regression from label-free DPC images.
- Hybrid architecture fuses convolutional and transformer features via a learnable cross-branch gating module.
- LLM generates biologically grounded summaries of predicted cell states.
- Experiments on BSCCM and Blood Cells Image benchmarks.
- Offers scalable, non-invasive alternative to fluorescence-based cytometry.
Entities
Institutions
- arXiv
- Berkeley Single Cell Computational Microscopy
- Blood Cells Image