MicroSSIM

MicroSSIM is an image measure aimed at addressing the shortcomings of the Structural Similarity Index Measure (SSIM), in particular in the context of microscopy images. Indeed, in microscopy, degraded images (e.g. lower signal to noise ratio) often have a different dynamic range than the original images. This can lead to a poor performance of SSIM.

The measure normalizes the images using background subtraction and a more appropriate range estimation. It then estimates a scaling factor used to scale the image to the target (original image or ground truth). The metric is then computed similarly to the SSIM.

MicroSSIM is easily extensible to other SSIM-like measures, such as Multi-Scale SSIM (MS-SSIM), for which we provide an example.

See the paper for more details.

Installation

pip install microssim

Usage

import numpy as np
from microssim import MicroSSIM, micro_structural_similarity
from skimage.metrics import structural_similarity

rng = np.random.default_rng(42)
N = 5
gt = 200 + rng.integers(0, 65535, (N, 256, 256)) # stack of different images
pred = rng.poisson(gt) / 10

# using the convenience function
result = micro_structural_similarity(gt, pred)
print(f"MicroSSIM: {result} (convenience function)")

# using the class allows fitting a large dataset, then scoring a subset
microssim = MicroSSIM()
microssim.fit(gt, pred) # fit the parameters

for i in range(N):
    score = microssim.score(gt[i], pred[i]) # score a single pair
    print(f"MicroSSIM ({i}): {score}")

# compare with SSIM from skimage
for i in range(N):
    score = structural_similarity(gt[i], pred[i], data_range=65535)
    print(f"SSIM ({i}): {score}")

The code is similar for MicroMS3IM.

Tips for deep learning

MicroSSIM was developed in the context of deep-learning, in which SSIM is often used as a measure to compare denoised and ground-truth images. The tips presented here are valid beyond deep-learning.

The larger the dataset, the better the estimate of the scaling factor will be. Therefore, it is recommended to fit the measure on the entire dataset (e.g. the whole training dataset). Once the data fitted, the MSSIM class has registered the parameters used for normalization and scaling. You can then score a subset of the data (e.g. the validation or test datasets) using the score method.

Cite us

If you use MicroSSIM in your research, please cite us:

Ashesh, Ashesh, Joran Deschamps, and Florian Jug. "MicroSSIM: Improved Structural Similarity for Comparing Microscopy Data." arXiv preprint arXiv:2408.08747 (2024). link.