A spectrum of diseases are characterized by high iron accumulation in various body tissues [1,2]. High iron overload can result from the disease itself (e.g. hereditary hemochromatosis) [3] or because of blood transfusions used as treatment for an underlying condition (e.g. thalassemia) [4,5]. In many of these cases, frequent monitoring of tissue iron concentration is required. In practice, an assessment of liver iron concentration (LIC) is often used as an estimate of total tissue iron storage. One of the well-known non-invasive LIC monitoring methods validated against liver biopsy is an R2-based MRI technique (“FerriScan”) developed by Resonance Health Limited (Burswood, WA, Australia) [6,7]. However, issues of cost, report turnaround time, availability, and administrative burden have motivated investigations into alternate methods of LIC quantification.

One alternate LIC quantification approach is R2* [[8], [9], [10]]. Previous studies have shown good correlation between LIC derived from R2* and FerriScan and/or biopsy [8,9,11,12]. However, during routine clinical use at our site, we observed that R2*-derived LIC sometimes underestimates FerriScan-derived LIC at very high values. A preliminary analysis suggested that the issue lay with insufficient SNR in one or more of the TEs used in the R2* fits. If SNR drops below a certain threshold, the noise statistics become non-Gaussian [13,14]. This can bias the resulting R2* value calculated from the decay data, thus reducing accuracy [[15], [16], [17]]. This phenomenon has been well studied, and a number of solutions proposed: The first approach fits the complex decay signal data (rather than the magnitude) to preserve the Gaussian nature of the noise statistics [10,[18], [19], [20]]; the second approach uses magnitude data, but models the non-Gaussian noise behaviour explicitly [16,17]; the last approach (truncation) removes data that lies below a specified noise threshold [15,17]. The advantage of the first two approaches is that all data is used. This potentially increases the precision of the R2* fits compared to the truncation method [10]. However, the first approach requires access to the complex signal data, as well as accounting for background phase effects, and the second approach requires knowledge of the underlying noise behaviour – information that is not always available in routine clinical practice or requires more dedicated pulse sequences. The challenge with truncation methods is deciding on the noise threshold. Existing methods aim to truncate data around the range where the noise statistics become significantly non-Gaussian. This approach throws away a minimum of data, minimizing the precision loss associated with truncation. The drawback with this approach is that if the noise magnitude is not known precisely, the truncation point may be incorrect. In fact, even if the noise magnitude is known precisely, there is likely a range of values rather than one distinct point over which noise behaviour transitions. This could lead to the inclusion of non-Gaussian data, reducing the accuracy of the R2* estimate. In particular, there could be a negative bias in the R2* estimate [15]. In this study, we take a different approach to truncation. Our objective is to choose a truncation threshold that minimizes the likelihood of including non-Gaussian data, rather than minimizing the amount of data thrown away. In effect, the conventional approach maximizes precision at a cost of accuracy, whereas our approach maximizes accuracy at a potential cost of precision. The technique will be validated against FerriScan-derived LIC. The hypothesis of this study is that the truncation-based method can provide accurate LIC estimates over the full range of LICs characterized by FerriScan.