ilex-hub/deep_dti.deepdti-nb1-ep100.1

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

from ilex.models.deep_dti import DeepDTI
model = 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.

Upstream license reference: https://opensource.org/licenses/MIT

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/DeepDTI

Provenance

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.