Biopsy-based calibration of T2* magnetic resonance for estimation of liver iron concentration and comparison with R2 Ferriscan

Abstract

Background: There is a need to standardise non-invasive measurements of liver iron concentrations (LIC) so clear inferences can be drawn about body iron levels that are associated with hepatic and extra-hepatic complications of iron overload. Since the first demonstration of an inverse relationship between biopsy LIC and liver magnetic resonance (MR) using a proof-of-concept T2* sequence, MR technology has advanced dramatically with a shorter minimum echo-time, closer inter-echo spacing and constant repetition time. These important advances allow more accurate calculation of liver T2* especially in patients with high LIC.

Methods: Here, we used an optimised liver T2* sequence calibrated against 50 liver biopsy samples on 25 patients with transfusional haemosiderosis using ordinary least squares linear regression, and assessed the method reproducibility in 96 scans over an LIC range up to 42 mg/g dry weight (dw) using Bland-Altman plots. Using mixed model linear regression we compared the new T2*-LIC with R2-LIC (Ferriscan) on 92 scans in 54 patients with transfusional haemosiderosis and examined method agreement using Bland-Altman approach.

Results: Strong linear correlation between ln(T2*) and ln(LIC) led to the calibration equation LIC = 31.94(T2*)-1.014. This yielded LIC values approximately 2.2 times higher than the proof-of-concept T2* method. Comparing this new T2*-LIC with the R2-LIC (Ferriscan) technique in 92 scans, we observed a close relationship between the two methods for values up to 10 mg/g dw, however the method agreement was poor.

Conclusions: New calibration of T2* against liver biopsy estimates LIC in a reproducible way, correcting the proof-of-concept calibration by 2.2 times. Due to poor agreement, both methods should be used separately to diagnose or rule out liver iron overload in patients with increased ferritin.

Figure 1

T2* method and inter-observer reproducibility. (A) A transverse slice of the liver where an example of the region of interest (ROI) is seen in green within the mid/lateral area free of large vessels. (B) T2* estimation using the truncation method in exponential curve fitting: crosses represent truncated data points whereby only the first 5 points contribute to the decay curve model. (C) The inter-observer reproducibility and agreement for R2* between observer 1 and 2 (both in triplicate, error bars show SEM, 96 scans) with line of identity. Coefficient of variation 5.79%. (D) Bland-Altman analysis of inter-observer liver R2* percentage differences plotted against the mean R2* of the two observers. The bias as the mean of all differences was 0.61 ± 4.29% (SD). The 95% limits of agreement (LoA) were between -7.79 and 9.02%.

Figure 2

The calibration of T2* and R2* models for LIC estimation. (A) The relationship of liver T2* to LIC obtained by biopsy in 50 samples. The regression fit from Figure  2C was plotted onto the data after exponentiation: LIC = 31.94(T2*)^-1.014 with 95% CI 27.8 to 36.7 (87-115%) for the first term and -1.118 to -0.91 (110-90%) for the exponent. (B) The relationship of liver R2* (=1000/T2*) to LIC of the data in 2A, with regression line from the model in 2D after exponentiation: LIC = 0.029R2*^1.014; 95% CI 0.016 to 0.054 (55-186%) for the first constant, 0.910 to 1.118 (90-110%) for the exponent. (C) Log-log plot of LIC versus liver T2*: Pearson r = -0.94 (95% CI -0.97 to 0.91, p < 0.0001); linear regression: ln(LIC) = 3.464-1.014ln(T2*); slope SE = 0.052, 95% CI -1.118 to -0.9 (110-90%); intercept SE = 0.069, 95% CI 3.325 to 3.603 (96-104%), r squared 0.89. (D) Log-log plot of LIC versus liver R2*: Pearson r = 0.94 (95% CI 0.90 to 0.97, p < 0.0001); linear regression: ln(LIC) = 1.014ln(R2*)-3.54; slope SE = 0.052, 95% CI 0.910 to 1.118 (90-110%); intercept SE = 0.30, 95% CI -4.152 to -2.928 (117-83%); r squared 0.89.

Figure 3

Relationship of new T2*-LIC to R2-LIC (Ferriscan) measurements. (A) R2-LIC plotted against T2*-LIC (derived from Equation 1) for comparison cohort; range with lowest scatter circled. Mixed model regression on whole range log-transformed data ln(R2-LIC) = 1.04×ln(T2*-LIC)-0.08, with slope and intercept 95% CI of 0.96 to 1.11 (p < 0.001) and -0.26 to 0.11 (p = ns), is shown here exponentiated to R2-LIC = 0.83×T2*-LIC^1.04 with 95% CI 0.96 to 1.11 and 0.55 to 1.29, r squared = 0.65. (B) Whole range Bland-Altman plot of percentage difference vs mean LIC with non-uniform scatter with linear regression correction. Corrected bias is -1.29*(mean LIC)+23.03; absolute residuals R were related to LIC: R = 0.47*(mean LIC)+17.44, p = 0.01, therefore, accounting for growth in variance, 95% LoA were from -0.14*(mean LIC)+65.9 to -2.44*(mean LIC)-19.8 or changing from ±45 to ±90% across the whole range. (C) Mixed model linear regression of R2-LIC on T2*-LIC for R2-LIC range 0–10 mg/g dw: R2-LIC = 0.87 × R2*LIC-0.55, slope 95% CI 0.74 to 0.99 (p < 0.001), intercept -0.01 to 1.19 (p = 0.089); r squared = 0.86. Insignificant intercept can be abandoned to give a proportionality with a slope of 0.96, 95% CI 0.89-1.02, p < 0.001, which being indistinguishable from 1, follows the line of identity. (D) Bland-Altman plot of the mean of the two methods (x-axis) plotted against their percentage difference (y-axis) for the range marked by circle in 3A, which for R2-LIC values <10 mg/g dw (here in 37 scans on 26 patients) shows non-uniform scatter corrected by linear regression. Corrected bias is -4.51*(mean LIC)+40.23, p = 0.007, SD of residuals 23.22%. Absolute residuals did not relate to LIC (p = 0.99), therefore 95%LoA were bias ± 1.96*23.22% or ±45.51%.

Figure 4

Slope comparison of R2* LIC calibration models. Slope comparison of R2* based LIC calibration methods (after J. Hankins et al. [18]): by L. Anderson LIC = 0.0146(R2*)-0.27 (green), by J. Wood LIC = 0.0254(R2*) + 0.202 (orange), by J. Hankins LIC = 0.028(R2*)-0.45 (blue), our method (RBH-UCLH) LIC = 0.032(R2*)-0.14 (red).