MedSimilarity

What is MedSimilarity?

MedSimilarity is an open source Python to compare 2D medical images.

Citation

If you use this software, please cite it in your publication using:

@software{medsimilarity,
  title={MedSimilarity},
  author={Kulkarni, Pranav},
  month={May},
  year={2023},
  url={https://github.com/UM2ii/MedSimilarity},
  doi={10.5281/zenodo.7937894}
}

Getting Started

MedSimilarity is currently not available through pip, but you can manually install it.

Manual Installation

You can manually install MedSimilarity as follows:

$ git clone https://github.com/UM2ii/MedSimilarity
$ pip install MedSimilarity/

Example Notebook

We have provided an example notebook in this repository, along with 100 test images to experiment with. You can find the example notebook here.

Documentation

medsimilarity.structural_similarity

Computes the mean structural similarity index measure (SSIM) between two images. This implementation is an extension of skimage.metrics.structural_similarity (https://scikit-image.org/docs/stable/api/skimage.metrics.html#skimage.metrics.structural_similarity) with preprocessing steps for medical images.

Arguments:

img1, img2: PIL.Image Input images

Returns:

score: float The mean structural similarity index measure over the image grad: ndarray The gradient of the structural similarity between img1 and img2 diff: ndarray The full SSIM image

Notes:

• Structural similarity is not invariant to transformations

medsimilarity.structural_comparison

Computes the pairwise structural similarity index measure (SSIM) between an image and a dataset and returns the top K matches.

Arguments:

img: str Path to image dataset: list List containing paths to each image in dataset top_k: int, optional Number of best matches for img in dataset use_multiprocessing: bool, optional Enables spawning of multiple processes to speed up pairwise SSIM calculation

Returns:

score: ndarray The top_k matches for img in dataset with SSIM score

medsimilarity.dense_vector_comparison

Computes the cosine similarity scores using dense vector representations (DVRS) between an image and dataset and returns the top K matches. This method uses SentenceTransformers ViT-B transformer for computation.

Arguments:

img: str Path to image dataset: list List containing paths to each image in dataset top_k: int, optional Number of best matches for img in dataset use_multiprocessing: bool, optional Enables encoding images into embeddings using multiprocessing. If device is 'cuda', images are encoded using multiple GPUs. If device is 'cpu', multiple CPUs are used device: str, optional Specifies device to move all resources to. Use 'cuda' to enable GPU acceleration. If left blank, by default 'cuda' is used if available. If not, 'cpu' is used

Returns:

score: ndarray The top_k matches for img in dataset with DVRS score

medsimilarity.combined_score

Experimental!

Computes the combined score from structural similarity index measure (SSIM) and dense vector representations (DVRS) scores for a pair of images using the formula:

x_combined = sqrt(x_ssim)*(x_dvrs)^2

Arguments:

x_ssim: float The SSIM score for pair of images x_dvrs: float The DVRS score for pair of images

Returns:

x_combined: float Combined score for pair of images

Notes:

• This worked well in my testing but please take this with a grain of salt!