PPSW-SHAP: Towards Interpretable Cell Classification Using Tree-Based SHAP Image Decomposition and Restoration for...

Area of Science:

• Biotechnology
• Medical Imaging
• Artificial Intelligence

Background:

• High-throughput microscopy is crucial for cell analytics in cell therapy (CT) manufacturing.
• Image noise and artifacts in microscopy can lead to inaccurate AI model findings.
• Current AI models struggle with noise without extensive training data.

Purpose of the Study:

• To develop a robust method for noise reduction in microscopy images for CT manufacturing.
• To create an interpretable machine learning (ML) platform for enhanced end-user understanding.
• To improve the precision of cell characterization in CT manufacturing.

Main Methods:

• Implemented a Periodic Plus Smooth Wavelet (PPSW) transform for image noise decomposition and restoration.
• Developed an interpretable ML platform utilizing tree-based Shapley Additive exPlanations (SHAP).
• Applied supervised clustering with mean SHAP values on decomposed bright-field images of human Mesenchymal Stem Cells (MSCs).

Main Results:

• The PPSW transform effectively reduced systematic noise in microscopy images.
• The SHAP-based ML platform provided end-to-end interpretability for AI models.
• Improved precision in cell characterization was achieved for MSCs under various culture conditions.

Conclusions:

• The combined noise reduction and interpretable ML approach enhances AI performance in CT manufacturing.
• This framework lowers the learning load on AI by addressing artifacts during pre-processing.
• The study demonstrates a significant advancement in reliable and understandable cell analytics for CT production.