LOTUS: Learning to Optimize Task-based US representations

Yordanka Velikova, Mohammad Farid Azampour, Walter Simson, Vanessa Gonzalez Duque, Nassir Navab

MICCAI 2023 (Oral Presentation)

arXiv Code (GitHub) Video

Abstract

Anatomical segmentation in ultrasound is critical for diagnosis and monitoring, but obtaining accurate pixel-wise annotations is both challenging and time consuming. LOTUS tackles this by learning intermediate ultrasound representations that are optimized directly for downstream tasks. Using annotated CT label maps as a simulation medium, we cast ultrasound propagation with a differentiable ray tracer, while simultaneously training segmentation and unsupervised domain adaptation networks. The renderer refines tissue parameters end-to-end, yielding optimized ultrasound simulations that better highlight target anatomy and generalize to real scans.

Presentation

MICCAI 2023 oral presentation recording.

Differentiable ultrasound renderer

LOTUS differentiable renderer pipeline — Ray-based renderer assigns ultrasound-specific parameters per tissue label to synthesize attenuation, reflection, and scatter maps.

Optimized image representations

LOTUS iteratively updates renderer parameters per organ-specific task, enhancing contrast for the targeted anatomy while damping distractors.

End-to-end pipeline

LOTUS end-to-end training pipeline — Differentiable rendering, domain adaptation, and segmentation training occur jointly to bridge the gap between simulated and real ultrasound.

Results

During inference, real ultrasound frames are translated into the optimized representation before segmentation.

Citation

If you build on this work, please cite:

@inproceedings{velikova2023lotus,
  title={LOTUS: Learning to Optimize Task-Based US Representations},
  author={Yordanka Velikova and Mohammad Farid Azampour and Walter Simson and Vanessa Gonzalez Duque and Nassir Navab},
  booktitle={Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  pages={435--445},
  year={2023}
}