gevi

Main files:

Anime.py : Is a class used for displaying 2d Movies from a 3D Matrix.

ImageAlignment.py: Is a class used for aligning two images (move and rotate as well as an unused scale option). This source code is based on https://github.com/matejak/imreg_dft . Is was ported to PyTorch such it can run in GPUs.

DataContailer.py: Main class for data pre-processing of the gevi raw data.

TODO / Known problems

Support for several part files was included but, due to missing data with several part files, never tested.

Installation

The code was tested on a Python 3.11.2 (Linux) with the following pip packages installed:

numpy scipy pandas flake8 pep8-naming black matplotlib seaborn ipython jupyterlab mypy dataclasses-json dataconf mat73 ipympl torch torchtext pywavelets scikit-image opencv-python scikit-learn tensorflow_datasets tensorboard tqdm argh sympy jsmin pybind11 pybind11-stubgen pigar asciichartpy torchvision torchaudio tensorflow natsort roipoly

Not all packages are necessary (probably these are enougth: torch torchaudio torchvision roipoly natsort numpy matplotlib) but this is our default in-house installation plus roipoly.

We used a RTX 3090 as test GPU.

For installing torch under Windows see here: https://pytorch.org/get-started/locally/

Data processing chain

SVD (requires donor and acceptor time series) remove_heartbeat: bool = True

• start automatic_load
  • try to load previous mask
  • start: cleaned_load_data
    • start: load_data
    • work on XXXX.npy
      • np.load
      • organize acceptor (move to GPU memory)
      • organize donor (move to GPU memory)
      • move axis (move the time axis of the tensor)
      • move intra timeseries
        • donor time series and donor reference image
        • acceptor time series and acceptor reference image
      • rotate inter timeseries
        • acceptor time series and donor reference image
      • move inter timeseries
        • acceptor time series and donor reference image
    • spatial pooling (i.e. 2d average pooling layer)
      • acceptor(x,y,t) = acceptor(x,y,t) / acceptor(x,y,t).mean(t) + 1
      • donor(x,y,t) = donor(x,y,t) / donor(x,y,t).mean(t) + 1
    • remove the heart beat via SVD from donor and acceptor
      • copy donor and acceptor and work on the copy with the SVD
      • remove the mean (over time)
      • use Cholesky whitening on data with SVD
      • scale the time series accoring the spatial whitening
      • average time series over the spatial dimension (which is the global heart beat)
      • use a normalized scalar product for getting spatial scaling factors
      • scale the heartbeat with the spatial scaling factors into donor_residuum and acceptor_residuum
      • store the heartbeat as well as substract it from the original donor and acceptor timeseries
    • remove mean from donor and acceptor timeseries (- mean over time)
    • remove linear trends from donor and acceptor timeseries (create a linear function and use a normalized scalar product for getting spatial scaling factors)
  • use the SVD heart beat for determining the scaling factors for donor and acceptor (heartbeat_scale)
    • apply bandpass donor_residuum (filtfilt)
    • apply bandpass acceptor_residuum (filtfilt)
    • a normalized scalar product is used to determine the scale factor scale(x,y) between donor_residuum(x,y,t) and acceptor_residuum(x,y,t)
    • calculate mask (optional) ; based on the heart beat power at the spatial positions
  • scale acceptor signal (heartbeat_scale_a(x,y) * result_a(x,y,t)) and donor signal (heartbeat_scale_d(x,y) * result_d(x,y,t))
    • heartbeat_scale_a = torch.sqrt(scale)
    • heartbeat_scale_d = 1.0 / (heartbeat_scale_a + 1e-20)
  • result(x,y,t) = 1.0 + result_a(x,y,t) - result_d(x,y,t)
  • update inital mask (optional)
• end automatic_load

Classic (requires donor, acceptor, volume, and oxygenation time series) remove_heartbeat: bool = False

• start automatic_load
  • try to load previous mask
  • start cleaned_load_data
    • start load_data
      • work on XXXX.npy
        • np.load (load one trial)
        • organize acceptor (move to GPU memory)
        • organize donor (move to GPU memory)
        • organize oxygenation (move to GPU memory)
        • organize volume (move to GPU memory)
        • move axis (move the time axis of the tensor)
        • move intra timeseries
          • donor time series and donor reference image; transformation also used on volume
          • acceptor time series and acceptor reference image; transformation also used on oxygenation
        • rotate inter timeseries
          • acceptor time series and donor reference image; transformation also used on volume
        • move inter timeseries
          • acceptor time series and donor reference image; transformation also used on volume
    • spatial pooling (i.e. 2d average pooling layer)
    • acceptor(x,y,t) = acceptor(x,y,t) / acceptor(x,y,t).mean(t) + 1
    • donor(x,y,t) = donor(x,y,t) / donor(x,y,t).mean(t) + 1
    • oxygenation(x,y,t) = oxygenation(x,y,t) / oxygenation(x,y,t).mean(t) + 1
    • volume(x,y,t) = volume(x,y,t) / volume(x,y,t).mean(t) + 1
    • frame shift
      • the first frame of donor and acceptor time series is dropped
      • the oxygenation and volume time series are interpolated between two frames (to compensate for the 5ms delay)
  • measure heart rate (measure_heartbeat_frequency) i.e. find the frequency f_HB(x,y) with the highest power in the frequency band in the volume signal
  • use "regression" (i.e. iterative non-orthogonal basis decomposition); remove offset, linear trend, oxygenation and volume timeseries
  • donor: measure heart beat spectral power (measure_heartbeat_power) f_HB(x,y) +/- 3Hz; results in power_d(x,y)
  • acceptor: measure heart beat spectral power (measure_heartbeat_power) f_HB(x,y) +/- 3Hz ; results in power_a(x,y)
  • scale acceptor and donor signals via the powers
    • scale(x,y) = power_d(x,y) / (power_a(x,y) + 1e-20)
    • heartbeat_scale_a = torch.sqrt(scale)
    • heartbeat_scale_d = 1.0 / (heartbeat_scale_a + 1e-20)
  • result(x,y,t) = 1.0 + result_a(x,y,t) - result_d(x,y,t)
• end automatic_load

DataContailer.py

Constructor

def __init__(
    self,
    path: str,
    device: torch.device,
    display_logging_messages: bool = False,
    save_logging_messages: bool = False,
) -> None:

automatic_load

def automatic_load(  
    self,
    experiment_id: int = 1,
    trial_id: int = 1,
    start_position: int = 0,
    start_position_coefficients: int = 100,
    fs: float = 100.0,
    use_regression: bool | None = False,
    # Heartbeat
    remove_heartbeat: bool = True,  # i.e. use SVD
    low_frequency: float = 5,  # Hz Butter Bandpass Heartbeat
    high_frequency: float = 15,  # Hz Butter Bandpass Heartbeat
    threshold: float | None = 0.5,  # For the mask
    # Extra exposed parameters:
    align: bool = True,
    iterations: int = 1,  # SVD iterations: Do not touch! Keep at 1
    lowrank_method: bool = True,
    lowrank_q: int = 6,
    remove_heartbeat_mean: bool = False,
    remove_heartbeat_linear: bool = False,
    bin_size: int = 4,
    do_frame_shift: bool = True,
    half_width_frequency_window: float = 3.0,  # Hz (on side ) measure_heartbeat_frequency
    mmap_mode: bool = True,
    initital_mask_name: str | None = None,
    initital_mask_update: bool = True,
    initital_mask_roi: bool = False,
    gaussian_blur_kernel_size: int | None = 3,
    gaussian_blur_sigma: float = 1.0,
    bin_size_post: int | None = None,
    calculate_amplitude: bool = False,
) -> tuple[torch.Tensor, torch.Tensor | None]: