Image quality and segmentation metrics

[jdeschamps]

A bunch of classes to compute image quality or segmentation metrics between two RAI. The results have been checked against original implementations (Cell Tracking Challenge, StarDist) or widely used librairies (scikit-image).

Image Quality:

• Mean squared error (MSE)
• Peak signal-to-noise ratio (PSNR)
• Structural similarity index measure (SSIM)

Segmentation:

• SEG (from the Cell Tracking Challenge)
• LazySEG: SEG calculation compatible with lazy loading and multithreading, accumulating the SEG score from each image pair added to the computation.
• MultiMetrics: TP, FP, FN, Precision, Recall, F1, Accuracy, mean matched and mean true IoU (inspired from StarDist)
• LazyMultiMetrics: MultiMetrics calculations compatible with lazy loading and multithreading, accumulating the SEG score from each image pair added to the computation.

[imagejan]

Thanks @jdeschamps for this work! I think some of the basic functionality should actually live in imglib2-roi where we have already similar methods related to overlapping labels.

[jdeschamps]

Right now this PR has to wait for the resolution of imglib/imglib2-roi#59 !

edit: imglib/imglib2-roi#59 has been merged

[jdeschamps]

New commits:

• Compatibility with the newly merged Add helper methods imglib2-roi#59
• Add normalized root mean square error (NRMSE) metrics

This PR is dependent on imglib2-roi 0.11.1-SNAPSHOT (currently 0.10.4 in the scijava-pom 30.0.0 and 0.11.0 in the scijava-pom 32.0.0-SNAPSHOT)

[ctrueden]

@jdeschamps Releasing imglib2-roi 0.12.0 now. Edit: It is released.

[jdeschamps]

@tpietzsch Any news on this front? :D

[tpietzsch]

@jdeschamps I rebased on current master (3dcd24c) and added a few easy simplifications. I will add more comments inline, but I think in principle this is fine as it is. Simplifications/optimizations can be made later.

The only real problem I see is MunkresKuhnAlgorithm (adapted from trackmate) which is under GPL license. This is incompatible with the BSD license of imglib2-algorithm. So either

• this class (and MultiMetrics, LazyMultiMetrics that use it) should be moved to imglib2-algorithm-gpl, or
• we should obtain permission to relicense it under BSD, or
• this should be swapped with a different, compatibly licensed, implementation of the hungarian algorithm.

Whatever you prefer.

[jdeschamps]

I get that!

I'd be in favor of relicensing. We should then ask @dscho whether he agrees to relicense MunkresKuhnAlgorithm.java from GPLv3 to BSD.

Do you know from the top of your head an alternative implementation? Otherwise I'd wait a bit and start looking for one in case @dscho disagrees.

[dscho]

I am fine with the proposed relicensing!

[tpietzsch]

Awesome! Thank you Johannes! @dscho

[tpietzsch]

finally...