Intensity diffraction tomography with a non-paraxial multiple-scattering model

Jiabei Zhu, Alex Matlock, Hao Wang, Lei TianBiophotonics Congress 2021 · OpenReview

TL;DR

This paper introduces a novel algorithm for high-fidelity 3D refractive index (RI) reconstruction from intensity-only microscopy images. By employing a non-paraxial multiple-scattering model (SSNP), it overcomes limitations of previous methods, enabling accurate imaging of complex biological samples with simple, fast microscopy setups.

What Problem Does This Paper Address?

Quantitative phase imaging (QPI) is crucial for visualizing biological samples in 3D by mapping their refractive index (RI). While optical diffraction tomography (ODT) is a common QPI technique, it requires complex interferometric setups. Intensity diffraction tomography (IDT) offers a simpler alternative, using only intensity measurements. However, accurately reconstructing the 3D RI distribution from IDT data, especially for strongly scattering biological samples and when using high-NA objectives (which capture more spatial frequencies but introduce non-paraxial effects), has been challenging. Existing multiple-scattering models often rely on approximations that limit their accuracy in these scenarios. This paper aims to bridge this gap by developing a more accurate and computationally efficient IDT reconstruction method.

Methods

The core of this work is the development and application of a Split-Step Non-Paraxial (SSNP) multiple-scattering model for IDT. Unlike previous models that often use paraxial approximations or local weak scattering assumptions, the SSNP model accounts for non-paraxial wave propagation and the coupling between diffraction and phase modulation more accurately.

Key methodological aspects include:

  • SSNP Forward Model: The SSNP model discretizes the sample into axial slices and propagates the wave field and its axial derivative slice-by-slice. This allows for a more faithful representation of scattering, especially at high angles.
  • Unified Reconstruction Algorithm: The authors developed a single algorithm that can handle both sequential (e.g., annular LED) and multiplexed (e.g., LED matrix) IDT acquisition strategies.
  • Automatic Differentiation: A memory-efficient automatic differentiation framework is employed to efficiently compute the gradients required for the inverse problem. This is crucial for handling large 3D datasets.
  • Regularized Least-Squares Optimization: The 3D RI distribution is reconstructed by solving a regularized least-squares optimization problem, utilizing a 3D total variation (TV) regularizer to suppress noise and enforce physical constraints.
  • GPU Acceleration: The algorithm is implemented using custom CUDA code for significant speedup on GPUs.

Datasets

The paper evaluates its method using both simulated data and experimental data from biological samples.

  • Simulated Data: A simulated 61-diameter bead was used to evaluate the forward model's accuracy under on-axis and off-axis (0.9 NA) illumination, comparing against Mie scattering theory. Further simulations involved spheres with different RI contrasts (Δn=0.01 and 0.05) to assess reconstruction quality.
  • Experimental Data:
    • Buccal epithelial cells: Reconstructed from annular IDT measurements (previously published).
  • Live C. elegans worm: Reconstructed from time-series annular IDT measurements.
  • Live C. elegans embryos: Reconstructed from multiplexed IDT measurements (previously published).

No new benchmark datasets were explicitly introduced, but the experimental datasets represent challenging biological samples for IDT.

Key Insights and Ideas

  • Non-paraxial SSNP is crucial for high-NA IDT: The SSNP model's ability to handle non-paraxial propagation significantly improves accuracy for high-NA illumination, where paraxial approximations fail.
  • Unified algorithm for diverse IDT setups: A single, efficient algorithm for both sequential and multiplexed IDT simplifies the reconstruction pipeline and broadens the applicability of the SSNP model.
  • Memory-efficient differentiation is key for 3D: The automatic differentiation approach, optimized for memory usage, is essential for reconstructing large 3D volumes on typical GPU hardware.
  • TV regularization effectively denoises: The use of 3D total variation regularization is vital for producing clean and sharp reconstructions, especially for noisy or complex samples.
  • Missing-cone artifacts remain a fundamental limit: Despite the improved model, the finite angular coverage of illumination and imaging optics still limits lateral resolution and introduces axial elongation.

Key Results

  • Forward Model Accuracy: Simulations showed the SSNP model accurately predicts intensity measurements for a bead under high-NA illumination, outperforming BPM and modified BPM models, especially for off-axis cases.
  • Reconstruction Quality:
    • Simulations demonstrated that the SSNP reconstruction algorithm recovers RI distributions with better accuracy than BPM and modified BPM, particularly for strongly scattering samples.
  • Experimental reconstructions of buccal epithelial cells showed high-resolution features and clear cell boundaries.
  • Reconstructions of live C. elegans worms and embryos showcased the ability to visualize complex internal structures and dynamic processes, with SSNP providing more robust demultiplexing and higher contrast than first-Born models.
  • Computational Performance: The GPU-accelerated SSNP-IDT algorithm achieves a significant speedup (30-60x compared to single-thread CPU) and can reconstruct a 600x600x150 voxel volume in approximately 120 seconds on a high-end GPU. The memory-efficient strategy reduces memory consumption by approximately 2x compared to a naive implementation.

Limitations and Future Work

  • Limited Fourier Coverage: The authors acknowledge that the limited angular coverage of their current IDT systems (up to 0.65 NA) leads to "missing-cone" artifacts, fundamentally limiting reconstruction quality, lateral resolution, and axial elongation.
  • Future Work:
    • Improved Fourier Coverage: Integrating annular and multiplexed illumination strategies to achieve better Fourier coverage and potentially higher NA optics.
  • Incorporating Dark-Field Information: Extending the SSNP model to handle dark-field measurements to capture high spatial frequency information beyond the objective NA.
  • Advanced Reconstruction Frameworks: Exploring integration with recent deep learning-based reconstruction methods (e.g., end-to-end supervised learning, deep denoisers, neural fields) to further suppress artifacts and improve reconstruction quality.
  • Optimizing GPU Implementation: Further optimization of the CUDA code for better utilization of GPU computing power.

Summary updated

BibTeX
@inproceedings{zhu2021intensity,
  author = {Jiabei Zhu and Alex Matlock and Hao Wang and Lei Tian},
  title = {{Intensity diffraction tomography with a non-paraxial multiple-scattering model}},
  booktitle = {Biophotonics Congress},
  year = {2021},
  url = {https://openreview.net/forum?id=VWdtSej3qj},
}
Figure