Machine Learning Based Lens-Free Shadow Imaging Technique for Field-Portable Cytometry.

Area of Science:

• Biophotonics
• Artificial Intelligence
• Biomedical Imaging

Background:

• Lens-free shadow imaging technique (LSIT) is cost-effective for microparticle and cell characterization.
• Current LSIT algorithms rely on handcrafted features, limiting adaptability to new cell types and struggling with noisy diffraction patterns.

Purpose of the Study:

• To develop an AI-powered signal enhancement and adaptive cell characterization method for LSIT.
• To overcome limitations of handcrafted features and improve classification accuracy and adaptability for diverse cell types.

Main Methods:

• Implemented a denoising autoencoder for signal enhancement.
• Utilized deep neural networks with transfer learning for adaptive cell characterization.
• Trained and validated the model on various cell types, including red blood cells (RBC) and white blood cells (WBC).

Main Results:

• Achieved signal enhancement of over 5 dB for cell diffraction patterns.
• Demonstrated classification accuracy exceeding 98% for tested cell types.
• The model successfully adapted to classify new cell types with minimal learning iterations.

Conclusions:

• The AI-driven approach significantly improves LSIT performance by enhancing signal quality and enabling robust, adaptive cell classification.
• This method offers a more versatile and accurate solution for automated cell analysis using LSIT, overcoming previous limitations.