openpros-inversion-net

InversionNet — OpenPros speed-of-sound reconstruction

Keras 3 port of the pretrained InversionNet baseline released with the OpenPros limited-view prostate ultrasound computed tomography (USCT) benchmark. It reconstructs a speed-of-sound map directly from simulated multi-source waveform data, without an iterative solver.

Because it is a Keras 3 model, it runs on the JAX, TensorFlow and PyTorch backends. The weights are converted from the original PyTorch checkpoint (model.pth, also in this repo) and reproduce its output to within float32 round-off (max abs. difference ~4e-5).

Usage

from zea.models.inversionnet import InversionNet

model = InversionNet.from_preset("inversionnet-openpros")
sos = model(waveforms)  # (B, 1000, 161, 40) -> (B, 401, 161, 1), in [-1, 1]

Inputs are waveforms of shape (batch, 1000 time samples, 161 receivers, 40 source/receiver combinations).

Preprocessing (required)

The weights only work on input scaled the way they were trained, and mis-scaled input degrades the reconstruction silently rather than raising:

1. Sign-preserving log compression, t(x) = sign(x) * log1p(|k * x|), with k = 1e5.
2. Min-max normalize [t(-0.25), t(0.45)] to [-1, 1], where -0.25 / 0.45 are the OpenPros data_min / data_max from its dataset_config.json.

The output is in [-1, 1] and maps linearly onto the OpenPros label range, 1300–3600 m/s.

k is part of the weights, not a tunable. The OpenPros job scripts pass --k 1e9, but k = 1e5 is what reproduces this checkpoint on the released data: on the OpenPros sample it gives a mean absolute error of 14 m/s against the ground-truth map, where k = 1e9 gives 244 m/s. The optimum is sharp — an order of magnitude either way costs roughly 3x in error.

See the zea documentation for the full pipeline.

Attribution

The architecture, the training data and these weights are the work of the OpenPros and InversionNet authors — please cite both when you use this model:

• Y. Wu and Y. Lin. InversionNet: An Efficient and Accurate Data-Driven Full Waveform Inversion. IEEE Transactions on Computational Imaging, 6:419–433, 2020. doi:10.1109/TCI.2019.2956866 · arXiv:1811.07875
• H. Wang, Y. Wu, Y. Feng, P. Jin, L. Zhang, S. Feng, J. Wiskin, B. Turkbey, P. A. Pinto, B. J. Wood, S. Luo, Y. Chen, E. Boctor and Y. Lin. OpenPros: A Large-Scale Dataset for Limited View Prostate Ultrasound Computed Tomography. 2025. arXiv:2505.12261 · code

Licensed CC-BY-4.0, following the original OpenPros release.

About

Converted and hosted as part of OpenH-RF, an open ultrasound channel-data initiative by NVIDIA, TU/e and Stanford, and distributed through zea.