# Ex-vivo validation of nine algorithms for quantifying infarcts with late gadolinium enhancement cardiovascular magnetic resonance

## TL;DR

This study compares nine algorithms for measuring heart infarct size using MRI against a detailed ex-vivo standard to determine their accuracy and reliability.

## Contribution

The study provides a direct ex-vivo validation of nine infarct quantification algorithms in a heterogeneous dataset, identifying which perform best.

## Key findings

- EWA, FWHM, and FACT showed accuracy and precision comparable to manual delineation.
- Heiberg-08 and FWHM are unsuitable for phase-sensitive inversion recovery images.
- Manual delineation by experienced observers remains a reliable method.

## Abstract

In cardiovascular magnetic resonance, late gadolinium enhancement (LGE) is the standard method to visualize myocardial infarction (MI). Many algorithms quantifying infarct size in LGE images exist. However, only few algorithms have been validated, i.e., benchmarked against an ex-vivo measurement. Furthermore, the reported algorithm performance varies considerably between studies.

The aim of this study was to compare the performance of all infarct measurement algorithms against an ex-vivo measurement and to promote a discourse regarding advantages and disadvantages of individual measurement methods.

MI was induced in 22 pigs. In-vivo LGE imaging was conducted on d0, d3 or d7 post-MI. For ex-vivo validation infarct was measured using high-resolution T1-weighted images. In-vivo infarct size was measured using the full-width at half-maximum (FWHM), n-SD from remote (2,3,5, and 6 SD), feature analysis and combined thresholding (FACT), expectation maximization-weighted A priori information (EWA), Heiberg-08 and Otsu algorithms and manual delineation. No manual adjustments were made to algorithm delineations.

Clear differences in variance and bias were observed between algorithm-based methods, and no method performed optimally in this heterogeneous dataset where the best had a bias of −0.48 ± 3.1, −0.3 ± 4.4%, 2.3 ± 4.2% left ventricle for EWA, FWHM, and FACT, respectively. Manual delineation by experienced observers performed well with a bias of 1.9 ± 5.4%.

EWA, Heiberg-08, FWHM, and FACT all perform on par with manual delineation, however, Heiberg-08, and FWHM are not suitable for phase sensitive inversion recovery images. The technique used to measure infarct size should be disclosed in clinical trials and in original research. Caution should be applied when comparing datasets employing different infarct quantification methods. Manual infarct delineation by experienced readers remains a reliable technique to measure infarct size.

Central illustration: Out of nine algorithms for infarct quantification on late gadolinium enhancement three had accuracy and precision on par with manual delineation when evaluated against high resolution T1w ex-vivo images. Estimated bias is indicated with a box and the precision is indicated by with bars and show the limits of agreement. The methods that performed acceptably were EWA (Expectation maximization Weighted A priori information), FWHM (Full Width Half Maximum), and FACT (Feature Analysis and Combined Thresholding).ga1

Central illustration: Out of nine algorithms for infarct quantification on late gadolinium enhancement three had accuracy and precision on par with manual delineation when evaluated against high resolution T1w ex-vivo images. Estimated bias is indicated with a box and the precision is indicated by with bars and show the limits of agreement. The methods that performed acceptably were EWA (Expectation maximization Weighted A priori information), FWHM (Full Width Half Maximum), and FACT (Feature Analysis and Combined Thresholding).