Tutorial

Code | Datasets

Software environment

The results reported in this paper were obtained using DeepPhD (version 1.1.1). The software environment includes Python 3.10, PyTorch 2.8.0, and CUDA 12.9. A step-by-step tutorial is available at https://cabooster.github.io/DeepPhD/Tutorial/. All relevant pretrained models are listed on the tutorial page. The core dependencies include TorchVision (version 0.23.0), Torchaudio (version 2.8.0), NumPy (version 2.2.6), SciPy (version 1.15.3), scikit-image (version 0.25.2), OpenCV-Python (version 4.12.0.88), pandas (version 2.3.3), Matplotlib (version 3.10.7), ImageIO (version 2.37.0), Pillow (version 11.3.0), einops (version 0.8.1), timm (version 1.0.20), FrEIA (version 0.2), TensorBoard (version 2.20.0), MMEngine (version 0.10.7).

The public repository currently specifies:

• Linux (recommended)
• Python 3.10
• NVIDIA GPU with CUDA 12.x
• A recent PyTorch build compatible with your GPU (select the matching CUDA wheel on pytorch.org)

Pinned remaining packages are listed in requirements.txt (install PyTorch first): numpy==1.26.4, scipy==1.11.4, scikit-image==0.22.0, tifffile==2024.8.30, imageio==2.34.2, pillow==10.4.0, tqdm==4.66.5.

Installation

git clone https://github.com/cabooster/DeepPhD.git
cd DeepPhD
conda create -n deepphd python=3.10 -y
conda activate deepphd

Install PyTorch first, matched to your CUDA version and GPU. Use the selector on pytorch.org to choose a build compatible with your driver and hardware (newer GPUs such as the RTX 5090 require a recent build with the appropriate architecture support). Example for CUDA 12.8:

pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 \
    --index-url https://download.pytorch.org/whl/cu128

Install the remaining dependencies:

pip install -r requirements.txt

Source code: https://github.com/cabooster/DeepPhD

Data format

Organize input volumes as multi-page TIFF stacks (.tif) in a single folder, for example:

your_dataset/
  ├── stack_001.tif
  └── stack_002.tif

Each TIFF should have shape T × H × W (time or depth × height × width). Stacks with fewer than 400 frames are automatically extended to meet the minimum length required for training.

All stacks in the same directory must come from the same imaging device (sensor), so that shared physical noise parameters (e.g., the FPN pattern and MPGN gain/variance) stay consistent within a single training or inference run. Do not combine data from different cameras or microscopes in one folder.

Noise model

Considering the dominant noise sources in fluorescence imaging, the overall noise model can be formulated as an additive combination of mixed Poisson–Gaussian noise (MPGN), fixed-pattern noise (FPN), and row noise (RN):

Please choose an appropriate noise model that matches how your data were acquired. The table below lists common recommendations:

Quote the --noise_model value in the shell ('fpn|rn|mpgn'), because | is a pipe operator.

Training

python DeepPhD_train.py \
  --exp_dir demo_lightsheet_zebrafish \
  --datasets_path /path/to/your_dataset \
  --noise_model 'fpn|rn|mpgn' \
  --save_noise

For multiphoton / PMT data, use --noise_model mpgn. By default, training runs on GPUs 0 and 1. To use different devices, pass --gpu (e.g., --gpu 0 or --gpu 0,1,2).

Key arguments:

Checkpoints are saved to:

results/<exp_dir>/saved_models/epoch_<N>.pth

Inference

python DeepPhD_inference.py \
  --exp_dir demo_lightsheet_zebrafish \
  --datasets_path /path/to/your_dataset \
  --noise_model 'fpn|rn|mpgn' \
  --save_noise

Denoised outputs (and optional noise maps) are saved under results/<exp_dir>/:

results/<exp_dir>/
  ├── saved_models/epoch_<N>.pth
  ├── plots/          # denoised TIFF stacks
  ├── FPN/            # learned FPN map, if --save_noise and the model includes FPN
  └── RN/             # estimated RN map, if --save_noise and the model includes RN