ZhenghanFang/DeepSTI-training-data

DeepSTI synthetic training data

This dataset contains the synthetic phantom arrays used by DeepSTI, the official implementation of DeepSTI: Towards Tensor Reconstruction using Fewer Orientations in Susceptibility Tensor Imaging (Medical Image Analysis, 2023).

The release preserves the NumPy directory layout consumed directly by the DeepSTI data loader. The subjects are organized as follows:

• Training: Sub001, Sub002, Sub007, Sub008, and Sub009
• Validation: Sub005
• Testing: Sub003 and Sub006

The training measurements are synthesized online from susceptibility tensor patches, masks, B0 directions, and volume metadata. Validation and test data include precomputed noisy phase measurements and dipole kernels at SNR 10 dB. The original unused partition phase cache is not included.

Download and train

From the root of a DeepSTI checkout:

hf download ZhenghanFang/DeepSTI-training-data \
  --repo-type dataset \
  --local-dir data/synthetic

# Optional: reproduce the patches from the included whole-image arrays.
# Existing patch files are left unchanged.
python scripts/generate_training_patches.py --data_dir data/synthetic

python deepsti/main.py \
  --mode train \
  --name default \
  --data_dir data/synthetic \
  --gpu 0

Citation

@article{fang2023deepsti,
  title={DeepSTI: Towards Tensor Reconstruction using Fewer Orientations in Susceptibility Tensor Imaging},
  author={Fang, Zhenghan and Lai, Kuo-Wei and van Zijl, Peter and Li, Xu and Sulam, Jeremias},
  journal={Medical Image Analysis},
  volume={87},
  pages={102829},
  year={2023},
  doi={10.1016/j.media.2023.102829}
}