OmniFHT: Pose-Free 3D Quantitative Phase Imaging of Flowing Cells

OmniFHT is a computational framework for pose-free 3D quantitative phase imaging (QPI) of flowing biological cells using Fourier Holographic Tomography (FHT).
It jointly estimates unknown 3D refractive-index (RI) distributions and cell poses directly from 2D holographic projections, without requiring controlled rotation or synchronization hardware.

The codebase is derived from and extends cryoDRGN with physics-aware forward models, modified network architectures, and customized training loops for FHT.

---

1. Features

• Physics-aware 3D reconstruction

  • Implements the Fourier Diffraction Theorem under the Rytov (weak scattering) approximation.
  • Uses an implicit neural representation (INR) to model a continuous 3D scattering potential.

• Pose-free joint inference

  • Simultaneous optimization of 3D volume, orientation (SO(3)), and in-plane translation.
  • Coarse-to-fine hierarchical pose search for robust initialization in complex rotational scenarios.

• Flexible data handling

  • Supports standard .mrcs stacks of holographic projections.
  • Test datasets are provided for quick end-to-end experiments.

• Analysis utilities

  • Scripts for pose analysis, FSC computation, and bubble/RI distribution diagnostics.

---

2. Repository Structure

At the top level, the repository is organized as:

OmniFHT/
├── OmniFHT/
│   ├── cryodrgn/              # Modified cryoDRGN core
│   │   ├── commands/          # Main CLI entry points (incl. abinit_homo.py)
│   │   └── ...                # Model, dataset, and training utilities
│   ├── analyze_pose_2D.py     # 2D pose analysis utilities
│   ├── analyze_pose_3D.py     # 3D pose / orientation analysis
│   ├── bubble_analyze.py      # Bubble / RI distribution diagnostics
│   ├── bubble_analyze_multi.py
│   ├── crop_center.py         # Cropping / centering helper
│   ├── fsc.py                 # FSC computation scripts
│   └── ...
├── test_data/                 # Example hologram stacks (.mrcs)
│   ├── 3_1.mrcs
│   ├── 4_1_64.mrcs
│   ├── 4_2_64.mrcs
│   ├── 4_3_64.mrcs
│   └── 7_1_64.mrcs
└── README.md

---

3. Installation

conda create -n omnifht python=3.9 -y
conda activate omnifht
pip install cryodrgn numpy scipy mrcfile tqdm matplotlib
git clone https://github.com/ai4imaging/OmniFHT.git
cd OmniFHT

---

4. Data Format

OmniFHT uses .mrcs stacks where each slice is a single holographic projection.

Example test data:

test_data/4_1_64.mrcs

---

5. Running Homogeneous Ab Initio Reconstruction

Example:

python OmniFHT/cryodrgn/commands/abinit_homo.py     test_data/4_1_64.mrcs     -b 4     -o runs/4_1_64_demo     -n 201     --checkpoint 100     --lr 0.01     --uninvert-data

For optical system settings, you can edit it in OmniFHT/cryodrgn/lattace.py line 29-31.

lambda_ = 0.5328
RI_ = 1.33
pixel_size = 5.86 / 40

---

6. Outputs

runs/4_1_64_demo/
│── weights.pkl
│── vol.mrc
│── poses.pkl
│── run.log
└── checkpoints/

7. Citation

If you use OmniFHT:

Pose-Free 3D Quantitative Phase Imaging of Flowing Cellular Populations, Ye et al., 2025. (In preparation)

---

8. Support

Please open GitHub Issues for bugs, questions, or feature requests.