Deep transfer learning-based hologram classification for molecular diagnostics.

Area of Science:

• Microscopy and Imaging Technologies
• Computational Biology
• Biomedical Diagnostics

Background:

• Traditional lens-based microscopy has limitations such as a restricted field of view.
• Lens-free digital in-line holography (LDIH) presents an alternative but requires extensive computation for image reconstruction.
• This computational demand hinders LDIH's application in resource-limited, point-of-care settings.

Purpose of the Study:

• To develop a computationally efficient method for processing LDIH images for cellular analysis.
• To enable rapid classification of cells using LDIH data without image reconstruction.
• To assess the feasibility of using deep transfer learning for point-of-care LDIH applications.

Main Methods:

• A deep transfer learning (DTL) approach was employed to process LDIH images.
• Holograms of cells labeled with molecular-specific microbeads were captured.
• Neural networks, including a VGG19 pretrained network, were trained to classify holograms directly from raw input data.

Main Results:

• The DTL-based approach successfully classified individual cells based on the number of bound microbeads using raw holograms.
• The trained neural networks demonstrated robust performance with experimental LDIH data.
• Reconstruction-free classification of cellular features was achieved.

Conclusions:

• Deep transfer learning provides an effective solution for computationally intensive LDIH image processing.
• This approach overcomes limitations of traditional LDIH, making it suitable for rapid cellular analysis.
• The integration of DTL with LDIH paves the way for low-cost, portable point-of-care diagnostic tools.