PathoS2VD

PathoS2VD generates a pseudo pathology z-stack from one 2D pathology image. It learns an 11-plane source z-axis prior, adapts it to a target image domain, and produces 33 planes at inference:

z00–z10: upper extension
z11–z21: central stack
z22–z32: lower extension

Install

Use Python 3.10 or 3.11:

pip install -r requirements.txt

Training and generation require access to the SVD-XT base model stabilityai/stable-video-diffusion-img2vid-xt and user-provided data and checkpoints. Configure paths with the PATHOS2VD_* environment variables used in the YAML files; no machine-specific paths are stored in this repository.

Data preparation

For raw source stacks arranged as:

SOURCE_ROOT/<slide>/z00...z18/<patch>.png

build the training annotation with:

python scripts/build_blur_motion_annotation.py --source-root SOURCE_ROOT

This writes SOURCE_ROOT/blur_motion_data6.csv, which is read directly by the source dataset loader.

Training and generation

accelerate launch scripts/train_source_vae.py --config configs/stage1/vae.yaml
accelerate launch scripts/train_stage1.py --config configs/stage1/diffusion.yaml
accelerate launch scripts/train_target_vae.py --config configs/stage2/aggc.yaml
accelerate launch scripts/train_stage2.py --config configs/stage2/aggc.yaml

python scripts/generate.py --config configs/inference/default.yaml

See configs/ for required environment variables and docs/ for evaluation and preprocessing details.

Acknowledgements

• The defocus-blur implementation is adapted from dfe-pr2010.

• The Stable Video Diffusion training implementation is adapted from SVD_Xtend.

• The VAE and diffusion components build on Hugging Face Diffusers.