Back to bookshelf

(TFUScapes / DeepTFUS) A Skull-Adaptive Framework for AI-Based 3D Transcranial Focused Ultrasound Simulation | Srivastav et al. 

science•research paper•1m•View Source
Appreciation
7
Importance
--
Date Added
5.11.26
TLDR
1) Provides TFUScapes, first large-scale high-res dataset of tFUS simulations through human skulls derived from MRIs. Each simulation returns a steady-state pressure field generated by a focused ultrasound transducer placed at realistic scalp locations. 2) Presents DeepTFUS, a U-Net that estimates normalized pressure fields directly from 3D CT volumes and transducer position.
2 Cents
Appreciate the 3D dataset release, but lacking important numbers for reproducing DeepTFUS, namely approximate parameter count (we don’t even know on what OOM).
Tags

  • Transducer is a single-element bowl (like a big contact lens). The sampled point cloud is to geometrically represent this transducer, which vibrates as a whole.

    • Their knobs for adjusment are only: radius, aperture, and position. (if you reran simulations you can change orientation and source amplitude, but they fix these).

  • U-Net is reasonable architecture here because we are mapping input to output of same resolution (one pressure field number per voxel), we need to incorporate both global information (hence downsampling) and local information (hence skip layers)

    • They have three branches to condition the U-Net on a transducer encoding z_T (PCA → Fourier positional encoding to improve representation? → MLP layers to what → Softmax and summation to what)
      1. Dynamic convolutions (following each regular conv) with grouped depth-wise kernels (does not mix channels) predicted from z_T put through an MLP
      2. Cross-attention
      3. FiLM (also does not mix channels) actually seems to hurt performance on max_pressure_error?

  • Notes on convolutions

    • The receptive field of each bottleneck voxel grows because each conv kernel reaches beyond a single voxel (kernel_size > 1, and downsampling compounds this growth by making each voxel at deeper layers stand for a larger chunk of original input.