patminton — GPU-Accelerated 3D Photoacoustic Tomography

PATminton: PAT Models ImplemeNTatiON.

Scalable GPU implementations of 3D photoacoustic tomography (PAT) models that account for the transducer spatial impulse response. The forward and adjoint operators are computed on the fly as matrix-vector products, so iterative image reconstruction never stores the system matrix. CGLS, L-BFGS-B, PGD and Chambolle-Pock TV solvers are built on top of them.

For a 200³ grid observed by 11 520 transducer positions with 1000 time samples, the dense system matrix would occupy 200³ × 11 520 × 1000 × 8 B ≈ 737 TB in double precision; patminton applies it and its adjoint on the fly on the GPU instead.

Installation

pip install patminton

Requirements: a CUDA-capable NVIDIA GPU (compute capability ≥ 7.0), an NVIDIA driver ≥ 525, and Python ≥ 3.9. On Linux x86-64 this installs a prebuilt wheel built with CUDA 12.8, which needs no CUDA toolkit and works with both CUDA 12 and CUDA 13 drivers and PyTorch builds. Its cuFFT dependency (libcufft.so.11) is installed from PyPI.

Installing from source instead requires the CUDA toolkit (nvcc ≥ 11 on PATH), which compiles the library for every supported architecture. To target a specific set of GPUs (e.g. V100 + A100 + RTX 30xx):

PATMINTON_CUDA_ARCH="70;80;86" pip install --no-binary patminton patminton

Quick example

import torch
from patminton import PAT, translation_rotation_system, least_squares_CG

infos = translation_rotation_system(
    transducer_radius=25e-3, transducer_height=7.5e-3,
    transducer_width=0.250e-3, transducer_pitch=0.298e-3,
    transducer_nbr_elements=64, transducer_wavelength=1500/5e6,
    grid_size=10e-3,
)

pat = PAT(200, 200, 200, 5e-3, 5e-3, 5e-3,
          nT=1024, tStart=12.5e-6, dt=16e-9, c=1500.0,
          mode='cylinder_lut', infos_transducers=infos)

p = torch.zeros((200, 200, 200), dtype=torch.float64, device='cuda')
p[100, 100, 100] = 1.0

s = pat @ p                     # forward:  signals from initial pressure
p_bp = pat.T @ s                # adjoint:  back-propagation

u, *_ = least_squares_CG(pat, s, M_inv=None, max_iter=50, lam=1e-4)

Documentation

patminton.readthedocs.io — installation, quickstart, physical and mathematical model, transducer models, reconstruction algorithms and key parameters.

To preview it locally:

pip install -r docs/requirements.txt
mkdocs serve      # live preview on http://127.0.0.1:8000

Tests

pip install .[test]
pytest              # CPU tests (no GPU needed)
pytest -m gpu       # operator tests on a CUDA GPU

Citation

If you use this package, please cite:

Trung-Thai Do, Paul Escande, Caroline Chaux, Jérôme Gateau, Hwee Kuan Lee — "Scalable implementations of photoacoustic tomography models accounting for transducers spatial impulse response", 2026.