Deep learning-based optical field screening for robust optical diffraction tomography.

Area of Science:

• * Computational imaging
• * Biophysics
• * Machine learning for scientific applications

Background:

• * Tomographic reconstruction quality is highly dependent on raw 2D image data integrity.
• * Manual inspection and rule-based methods for screening defective data are inefficient and inaccurate.
• * Robust quality control is crucial for high-throughput optical diffraction tomography (ODT).

Purpose of the Study:

• * To develop a deep learning-enabled quality control system for holographic data in ODT.
• * To improve the accuracy and throughput of screening defective 2D images.
• * To enhance the quality of reconstructed tomograms.

Main Methods:

• * A deep convolutional neural network was trained using an extensive database of annotated optical field images (clean/noisy).
• * The network was trained for binary classification of image quality.
• * Model interpretability was assessed using saliency maps.

Main Results:

• * The deep learning model achieved >90% test accuracy, outperforming non-expert visual inspection and a rule-based algorithm.
• * The superior screening performance led to a significant improvement in tomogram quality.
• * The model demonstrated generalizability on unseen biological cell data from a different experimental setup.

Conclusions:

• * Deep learning offers a robust and high-throughput solution for quality control in ODT.
• * The proposed network can serve as an effective module within tomographic reconstruction pipelines.
• * AI-based quality control enhances the reliability and efficiency of ODT for biological imaging.