ESPIRiT

PyTorch-based ESPIRiT coil sensitivity calibration for MRI.

Single codebase that runs on CPU, CUDA GPU, and Apple Silicon (MPS) — no separate code paths needed.

Notice

This package contains a PyTorch translation of the ESPIRiT implementation from the BART (Berkeley Advanced Reconstruction Toolbox), © 2013–2026 The Regents of the University of California and BART Developer Team. BART is licensed under the BSD 3-Clause License. See https://codeberg.org/mrirecon/bart.

Usage

1) CLI

uvx espirit kspace.npy

2) Python

uv add espirit

import numpy as np
import torch
from espirit import espirit

# NumPy
kspace_np = np.load("kspace.npy")
csm_np = espirit(kspace_np)

# PyTorch 
kspace_pt = torch.randn(8, 24, 24, 24, dtype=torch.complex64)
csm_pt = espirit(kspace_pt)

Options

csm = espirit(
    kspace,             # (n_coils, *spatial_dims)
    calib_size=24,      # calibration region size
    kernel_size=6,      # sliding-window kernel size
    threshold=0.001,     # singular-value threshold
    mask_threshold=0.8, # eigenvalue mask threshold
    normalize=True,     # RSS=1 normalization
    rotphase=True,      # remove phase ambiguity
    device=None         # cuda, mps, or cpu (auto-detect when None)
)

Device support

Device	Backend	Notes
cpu	NumPy/MKL	Always available
cuda	NVIDIA GPU	Requires CUDA toolkit
mps	Apple Metal	macOS with Apple Silicon

The same code runs on all devices — PyTorch handles dispatch automatically.