Skip to content

Latest commit

 

History

History

fuseimg

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
data
data
 
 
datasets
datasets
 
 
eval
eval
 
 
utils
utils
 
 
README.md
README.md
 
 
__init__.py
__init__.py
 
 
requirements.txt
requirements.txt
 
 

README.md

FuseMedML - extension for imaging

Data package

Extends FuseMedML data package for imaging. Here you can find useful ops for imaging and public datasets implementations

Imaging operators

color

  • OpClip - clip values pixels/voxels values, support both torch tensor and numpy array
  • OpNormalizeAgainstSelf - normalizes a tensor into [0.0, 1.0] using its img.min() and img.max() (NOT against a dataset)
  • OpToIntImageSpace - normalizes a tensor into [0, 255] int gray-scale using img.min() and img.max() (NOT against a dataset)
  • OpToRange - linearly project from a range to a different range

image loader

  • OpLoadImage - loads variety of medical imaging formats from disk

shape

  • OpHWCToCHW - transform HWC (height, width, channel) to CHW (channel, height, width)
  • OpCHWToHWC - transform CHW (channel, height, width) to HWC (height, width, channel)
  • OpSelectSlice - select one slice from the input tensor,
  • OpFindBiggestNonEmptyBbox2D - Finds the the biggest connected component bounding box in the image that is non empty (dark)
  • OpFlipBrightSideOnLeft2D - Returns an image where the brigheter half side is on the left, flips the image if the condition does nt hold.
  • OpResizeAndPad2D - Resize and Pad a 2D image
  • OpSelectSlice - select one slice from the input tensor
  • OpPad - pad a given image on all the sides

augmentation

color

  • OpAugColor - color augmentation for gray scale images of any dimensions, including addition, multiplication, gamma and contrast adjusting
  • OpAugGaussian - add gaussian noise to numpy array or torch tensor of any dimensions

geometry

  • OpAugAffine2D - 2D affine transformation for torch tensors
  • OpAugCropAndResize2D - alternative to rescaling in OpAugAffine2D: center crop and resize back to the original dimensions. if scale is bigger than 1.0, the image is first padded.
  • OpAugSqueeze3Dto2D - squeeze selected axis of volume image into channel dimension, in order to fit the 2D augmentation functions
  • OpAugUnsqueeze3DFrom2D - unsqueeze back to 3D, after you apply the required 2D operations
  • OpResizeTo - resize an image into given dimensions
  • OpCrop3D - crop 3d image to certain size. can be used as random crop using OpSample.
  • OpRotation3D - rotate 3d image across the 3 planes xyz.

debug

  • OpVis2DImage - Visualize a 2D image (either display it or save it as an image).
  • OpVisImageHist - Visualize single image histogram (either display it or save it as an image).

Imaging datasets

  • kits21 - 2021 Kidney and Kidney Tumor Segmentation Challenge Dataset. See link
  • stoic21 - Dataset created for COVID-19 AI challenge. Aims to predict the severe outcome of COVID-19, based on the largest dataset of Computed Tomography (CT) images of COVID-19
  • cmmd - dataset that contains breast mammography biopsy info and metadata from chinese patients - https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=70230508. Aims to predict breast cancer biopsy results from mammography
  • mnist - mnist dataset implementation
  • isic19 - 2019 International Skin Imaging Collaboration (ISIC) Challenge dataset. The goal is to classify demoscropic images among nine different diagnostic categories
  • knight - KNIGHT Challenge dataset. The goal is to classify cases into risk groups based on CT image and clinical data. See link