Editorial to “Enhancing Study Design and Analysis of MR Imaging Markers Through Measurement Error Modeling”

Similar to molecular markers, quantitative imaging biomarkers require high measurement precision and reproducibility to fully realize their capabilities in precision medicine, from improving disease diagnosis to assessing treatment response and guiding therapeutic decisions. Measurement errors, including systematic bias and random error from biological variation [1, 2], pose a fundamental challenge in quantitative imaging. These errors can arise due to multiple factors, such as differences in readers, scanners, imaging protocols, acquisition conditions, and computational algorithms. While these issues exist in single-center studies, they are amplified in multi-site studies, where pooling data across institutions introduces additional variability despite yielding larger sample size and greater statistical power. Failure to account for them can bias effect estimation and reduce power to detect associations, and lead to underestimation of the marker's clinical utility [3]. Therefore, understanding the source of variability and addressing measurement error is essential to ensure the validity of the findings, improve reproducibility, and ultimately enhance patient care. Various statistical methods have been developed to mitigate measurement error or batch effect in imaging marker evaluations, many of which are adapted from approaches originally designed for genomic data. These methods can be broadly categorized into three classes: location–scale adjustment, matrix factorization, and discretization [4]. ComBAT is one of the most widely adopted approaches for multi-site imaging harmonization [5-7]. By estimating and correcting site effects in the acquired measurements, ComBAT reduces inter-site variability and improves the alignment of marker distributions across sites [2]. Several extensions have also been recently developed with various enhancements [8, 9]. In this issue, Ma et al. propose a dual data collection design to adjust for measurement error in MRI markers [10]. In addition to the main study of imaging data from the population of interest, they collect a reliability dataset consisting of scan-rescan acquisition from a small number of healthy volunteers on different scanners from presenting sites, which is used to quantify within-subject variability of the MRI measurement of interest. Using the resulting reliability coefficients, their method performs regression calibration on the main study dataset to recover measurement error-corrected true values, thereby enabling unbiased estimation of associations between imaging biomarkers and outcomes of interest. In simulation studies under small site effects, this dual data collection design produces unbiased estimations. It can also be readily integrated with ComBAT harmonization, offering improved estimation compared with using either ComBAT or the proposed measurement error correction alone in scenarios with large magnitude of side effects bias. Moreover, this dual design facilitates effective planning of the study with a formula for sample size calculation that explicitly accounts for the measurement error across sites. The method ensures adequate statistical power, particularly when image marker measurements have limited reproducibility, as shown in the case study of MRI T1w marker. Several considerations should be noted when applying this method. Both the reliability study and the main study require careful planning. The accurate measurement error adjustment relies on the reliability coefficient, which must be estimated from an appropriate reliability dataset. The generalizability of such coefficient also remains to be explored. Because the reliability coefficient is sensitive to the variance and the spread of marker values, the participants and tissue type in the reliability dataset should be aligned with the population in the main study to ensure the regression calibration is based on compatible reliability estimates. In this paper, the authors estimated the reliability coefficient for MRF T1 and MRI T1w on three different scanners. While the difference in estimated coefficients of the two biomarkers suggests they are marker-specific, whether the coefficient is scanner-specific or not is unknown. Given the many possible sources of measurement error during image acquisition, processing and analysis, the reliability coefficient should be used with caution or re-estimated when applied in a different setting. Future research could extend the method to address complex measurement errors and assess how various factors impact its applicability. Nevertheless, the dual data collection design increases the statistical power under scenarios with a large site effect and informs the sample size estimation through properly accounting for measurement error. It provides a useful tool that strengthens the design and analysis in MRI biomarker research. This work was supported by NIH/NCI (P30CA006973).