Generalizable self-supervised learning for imaging flow cytometry on multi-dataset leukocyte differential.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Hematology

Background:

• Supervised learning in imaging flow cytometry (IFC) offers high leukocyte classification accuracy.
• However, the need for annotated cell images and associated cell loss limits its use for leukocyte differential analysis.
• Current IFC methods struggle with differential analysis due to labeling inefficiencies.

Purpose of the Study:

• To develop a self-supervised learning framework for accurate leukocyte differential analysis using IFC.
• To overcome the limitations of supervised learning, specifically the requirement for annotated data and cell loss.
• To enable high-accuracy leukocyte differential analysis using non-annotated images.

Main Methods:

• A self-supervised contrastive learning framework was pretrained on a non-annotated dataset of bright-field leukocyte images from custom IFC.
• A frozen feature extractor was utilized with lightweight classification heads (linear or MLP) for downstream tasks.
• The method was evaluated on three independent annotated datasets: 4-class normal subtypes, 2-class leukemia cells, and 8-class mixed normal/abnormal cells.

Main Results:

• The proposed method achieved high classification accuracies: 96.16% (4-class), 96.68% (2-class), and 92.24% (8-class).
• Performance was comparable to supervised baselines, with accuracy differences less than 2%.
• Strong generalization ability was demonstrated with F1-scores of 96.68% (2-class) and 91.92% (8-class) on unseen abnormal cell types.

Conclusions:

• The self-supervised learning framework enables high-accuracy leukocyte differential analysis from non-annotated IFC images.
• The method demonstrates strong generalization capabilities, even with abnormal cell types not present during pre-training.
• This approach overcomes the data annotation bottleneck and cell loss issues inherent in supervised IFC methods.