DeepDTI -- physics-informed 6-direction diffusion tensor imaging via 3D DnCNN -- DeepDTI 7-channel DnCNN (epoch 100)
Description
DeepDTI (Tian et al., NeuroImage 2020) is a physics-informed
deep-learning pipeline for high-fidelity 6-direction diffusion
tensor imaging. The model is a 3D DnCNN residual denoiser that
maps a 7-channel input (1 b=0 image + 6 DWIs along optimised
diffusion-encoding directions that minimise the condition
number of the diffusion tensor transformation matrix) to a
7-channel residual; subtracting the residual from the input
produces the denoised volumes that downstream consumers feed to
standard diffusion tensor fitting (FSL dtifit / MRtrix3
tckgen).
Architecture: 10-layer 3D DnCNN with 128 filters per intermediate
layer. Conv1 (7->128, ReLU); 8 middle Conv+BatchNorm+ReLU blocks
(128->128 each); Conv10 (128->7, linear). The BatchNorm running
stats are baked in at extract time and never updated at inference.
v0 ships the published deepdti_nb1_ep100.h5 checkpoint
(3,592,583 scalars total: 3,590,535 trainable + 2,048 BN running
stats).
Intended use
Map a 7-channel diffusion-tensor input (1 b=0 + 6 DWIs) to a 7-channel residual; subtract from input to obtain denoised volumes for downstream tensor fitting and tractography. Forward is shape-agnostic; the upstream trained at (64, 64, 64) block size.
Usage
1from ilex.models.deep_dti import DeepDTI
2model = DeepDTI.from_pretrained('ilex-hub/deep_dti.deepdti-nb1-ep100.1')
Authors
Tian Q., Bilgic B., Fan Q., Liao C., Ngamsombat C., Hu Y., Witzel T., Setsompop K., Polimeni J. R., Huang S. Y. (Massachusetts General Hospital, Harvard Medical School)
Citation
Tian Q., Bilgic B., Fan Q., Liao C., Ngamsombat C., Hu Y., Witzel T., Setsompop K., Polimeni J. R., Huang S. Y. (2020). DeepDTI -- High-fidelity Six-direction Diffusion Tensor Imaging using Deep Learning. NeuroImage 219, 117017. doi:10.1016/j.neuroimage.2020.117017.
References
- Tian Q., Bilgic B., Fan Q., Liao C., Ngamsombat C., Hu Y., Witzel T., Setsompop K., Polimeni J. R., Huang S. Y. (2020). DeepDTI -- High-fidelity Six-direction Diffusion Tensor Imaging using Deep Learning. NeuroImage 219, 117017. doi 10.1016/j.neuroimage.2020.117017.
- Tian Q., Li Z., Fan Q., Ngamsombat C., Hu Y., Liao C., Wang F., Setsompop K., Polimeni J. R., Bilgic B., Huang S. Y. (2021). SRDTI -- Deep learning-based super-resolution for diffusion tensor MRI. arXiv 2102.09069.
- Upstream code + weights -- github.com/qiyuantian/DeepDTI (MIT).
License
HF Hub license tag: mit
Effective terms: MIT (copyright (c) 2021 Qiyuan Tian). Network code + pretrained .h5 are both MIT-licensed (github.com/qiyuantian/DeepDTI). The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.
Copyright
DeepDTI is copyright (c) 2021 Qiyuan Tian (qiyuantian, Harvard).
MIT-licensed on both the network code
(github.com/qiyuantian/DeepDTI) and the released .h5 checkpoint.
The ilex JAX / Equinox port code is separately licensed under
Apache-2.0 / GPL-3.0.
Upstream source
Original weights / reference implementation:
https://github.com/qiyuantian/DeepDTIProvenance
This artefact was produced by
ilex's
save/load pipeline. The architecture is implemented in
ilex.models.deep_dti.DeepDTI and the weights have been converted
from their upstream format. See the upstream source above
for the canonical reference.
