https://github.com/buswinka/lsd_pure_torch

Local Shape Descriptors For Neuron Segmentation, but only pytorch.
https://github.com/buswinka/lsd_pure_torch

Last synced: 21 days ago
JSON representation

Local Shape Descriptors For Neuron Segmentation, but only pytorch.

• Host: GitHub
• URL: https://github.com/buswinka/lsd_pure_torch
• Owner: buswinka
• License: mit
• Created: 2023-03-16T15:54:05.000Z (almost 2 years ago)
• Default Branch: main
• Last Pushed: 2024-02-14T17:10:24.000Z (11 months ago)
• Last Synced: 2024-11-06T17:57:23.235Z (2 months ago)
• Language: Python
• Size: 43 KB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# lsd_pure_torch

## Local Shape Descriptors For Neuron Segmentation, but only pytorch. 

This is more or less a rewrite of the work of Arlo Sheridan published here: https://www.nature.com/articles/s41592-022-01711-z

However removing numpy as a dependency. All functions should work with pytorch arrays, be differentiable, and cuda

compatible. 

Additionally, there should be a minimally functional training setup in scr/train/engine.py

This was originally written to be used in the Biomedical Image Segmentation Models (BISM) library: https://github.com/buswinka/bism

To calculate LSD of a 3D instance mask

```python

from lsd_pure_torch import lsd

import torch

import numpy as np

import skimage.io as io

from typing import Tuple

instance_mask: np.ndarray = io.imread('path/to/your/image.tif')  ## shape: [X, Y, Z], dtype: np.uint32

instance_mask: torch.Tensor = torch.from_numpy(instance_mask)

instance_mask = instance_mask.to('cuda').unsqueeze(0) ## shape: [C=1, X, Y, Z], dtype: np.uint32, device='cuda'

sigma: tuple[float, float, float] = (8., 8., 8.)  # Standard deviation affecting shape descriptors. 

voxel_size: tuple[int, int, int] = (1, 1, 5)  ## Relative voxel anisotropy. In this case, Z is 5 time spatially larger than X and Y

# LSD will return a 4D Tensor with shape [C=10, X, Y, Z], dtype=torch.float, device=instance_mask.device

local_shape_descriptors: torch.Tensor = lsd(segmentation=instance_mask, sigma=sigma, voxel_size=voxel_size)  

```