Interpretable unsupervised learning enables accurate clustering with high-throughput imaging flow cytometry.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning

Background:

• High-throughput imaging flow cytometry (IFC) generates vast datasets, posing challenges for analysis, especially when ground truth labels are unavailable.
• Unsupervised learning methods are needed to extract meaningful information from label-free IFC data without prior biological knowledge.

Purpose of the Study:

• To develop and evaluate an unsupervised deep embedding algorithm for clustering label-free IFC images.
• To demonstrate the model's capability in identifying distinct cell populations, including those with subtle or non-human-recognizable features.

Main Methods:

• Implementation of a Deep Convolutional Autoencoder-based Clustering (DCAEC) model for unsupervised learning.
• Encoding of IFC images into latent representations for subsequent clustering.
• Utilizing gradient-weighted class activation mapping (Grad-CAM) for interpreting model-identified features.

Main Results:

• Achieved high balanced accuracy for human white blood cell (WBC) clustering (91.9%) and WBC/leukemia classification (97.9%) using 3D IFC data.
• Demonstrated successful clustering (85.3% balanced accuracy) from label-free 2D transmission and 3D side scattering images, revealing non-human-recognizable patterns.
• Gradient-weighted class activation mapping identified salient, cluster-specific visual patterns, aiding in the interpretation of neural network feature recognition.

Conclusions:

• The DCAEC model provides an effective unsupervised approach for analyzing complex, label-free IFC data.
• The method enables the discovery of biologically relevant cell clusters, even when features are not visually apparent to humans.
• This work represents a significant step towards interpretable deep learning for high-throughput cell analysis using IFC.