Introduction

The harmonization of data is becoming increasingly important in imaging research. In particular, imaging data is affected by technical variability between scanners, which complicates comparisons across imaging sites, different scanners, and time points [1]. This issue impacts widely used cross-sectional modalities like magnetic resonance imaging (MRI), diffusion tensor imaging (DTI), and computed tomography (CT), as well as derived measurements such as region of interest (ROI) volumes, regional analysis of volumes examined in normalized space maps, cortical thickness estimates, and connectome matrices. To enhance statistical power when aggregating data from multiple sources, post-processing harmonization techniques are essential for reducing unwanted variability [2].

A single-source dataset introduces the potential for institutional biases, which may affect the generalizability of the model. To address the problem, large-scale, multi-center studies were utilized, but the integration of data from diverse sources is essential for achieving robust and generalizable findings. Harmonizing and disseminating data across sites enables researchers to capture variability across populations, imaging systems, and clinical practices, thereby strengthening the study’s validity. However, this process requires a well-designed infrastructure capable of acquiring, processing, and sharing data from multiple modalities while aligning with the workflows of all participating centers [3].

Sources of variability in quantitative imaging

The reliability of MRI-derived measurements of human cerebral cortical thickness was investigated by Han et al. [4], with a focus on the effects of field strength, scanner upgrades, and manufacturer differences. They found that the average variability was 0.15 mm for cross-scanner comparisons (Siemens/GE) and 0.17 mm for cross-field strength comparisons (1.5 T/3 T). Measurements across field strengths showed a slight bias, with cortical thickness appearing greater at 3 T [4]. In the variability assessment of volumetric data combined from five different scanners (2 General Electric Signa, 2 Siemens Symphony, and a Philips Gyroscan) at five different sites by repeating the scans of five volunteers at each of the sites using T1-only acquisitions, the two Siemens scanners exhibited a characteristic bias, overestimating white matter and underestimating gray matter compared to the other scanners. This bias, however, was not apparent when multimodal data were used [5]. The results indicated that the greatest compatibility between scanners is achieved when using equipment from the same manufacturer and maintaining image acquisition parameters as similar as possible.

Many factors can influence quantitative measurements during image acquisition. For example, the ability to select among various acquisition parameters and establish optimized protocols contributes to the diversification of positron emission tomography-CT (PET-CT) imaging techniques. Modern scanners incorporate numerous components that may affect quantitative accuracy, including 3D acquisition schemes, scintillators with intrinsic radioactivity, iterative reconstruction algorithms, CT-based attenuation correction, and scatter correction models that rely on multiple assumptions [6]. Moreover, the injected dose of [¹⁸F]Fluorodeoxyglucose may range from 300 to 700 megabecquerels, depending on scanner-specific characteristics such as the PET-CT detector material and acquisition mode [7]. These factors in acquisition may lead to research and clinical outcomes. To better understand and control such variability, long half-life PET-CT calibration phantoms are used to compare quantitative measurements across scanners, acquisition protocols, and processing methods by eliminating patient-related factors. Studies using such phantoms have demonstrated that PET-CT measurements exhibit both variance and size-dependent bias influenced by object dimensions, ROI definition, scan duration, acquisition mode, and reconstruction parameters, with appreciable biases reported even for relatively large (37 mm) objects [8, 9].

Snaith et al. [10] reported substantial variation in pelvis radiography techniques, with corresponding implications for clinical decision-making. Calls for standardization of pelvis radiographic studies have been made, and some authors have even proposed specific acquisition protocols [11]. However, there is no evidence that these protocols have been implemented in diagnostic imaging centers [10].

Reconstruction algorithms differ across manufacturers and software platforms. An et al. [12] compared two of the latest 3D modeling software packages, Syngo and Mimics, for accuracy and computational efficiency. Using CT scan images in DICOM (Digital Imaging and Communications in Medicine) format, they evaluated segmentation accuracy, anatomical measurements, cost, and computational time as benchmarks. The authors reported that Mimics outperformed Syngo in terms of semi-automated segmentation and equipment cost, whereas Syngo demonstrated superior computational efficiency [13]. Another key challenge in the reconstruction algorithm process is the potential for intensity bias in the slice data, often caused by anatomical motion relative to the imaging coils. Consequently, slices capturing the same anatomical region at different times may display varying sensitivity. Such bias field inconsistencies can introduce artifacts into the final 3D reconstruction, affecting both the clinical interpretation of critical tissue boundaries and the automated analysis of the data [14].

Beyond scanners, acquisition protocols, and processing methods, variability can also arise from the human element. Carapella et al. [15] demonstrated that standardized training of operators performing manual post-processing of cardiac MRI T1 maps improved consistency in the quantification of T1 biomarkers by reducing subjective bias. Training led to more accurate estimation of mean left ventricular myocardial T1 values and wall thickness, reduced variability in these measurements, and decreased discrepancies relative to reference standards. Moreover, patient positioning performed by technologists has also been shown to significantly influence both radiation dose and image quality in CT [16].

Lange et al. [17] aimed to assess the inter-study reproducibility of cardiac MRI cine image-based hemodynamic forces (HDF) measurements and to explore the current capabilities and limitations of this emerging deformation imaging technique. They concluded that inter-study variability could be improved through further software optimization, and emphasized the need for additional validation studies to support the broader clinical adoption of cardiac MRI-based HDF analysis [17]. Respectively, the study of 25 athletes’ cardiac MRI was analyzed by two independent observers and then re-analyzed by the same observer one week apart using HDF parameters derived from feature-tracking cardiac MRI showed a low inter- and intra-observer variability [18].

Strategies for harmonization

Recommended workflow for harmonized imaging epidemiology

Data harmonization can be conducted either retrospectively or prospectively (Figure 1). In both approaches, the first step for researchers is to identify the variables to be harmonized. This decision should be guided by the overarching goal of the harmonization effort, whether it is theory-driven (aimed at testing specific relationships among selected variables) or data-driven (focused on exploring relationships across a broader set of variables). Additionally, the availability of data and the acceptable degree of harmonization must also be taken into account [32].

Display full size

Figure 1. Comprehensive workflow for harmonizing multicenter quantitative imaging datasets. The process comprises three stages: (1) Preparation, in which study goals, variables, and harmonization pathways are defined; (2) Implementation, which includes retrospective and prospective harmonization procedures such as data standardization, intensity normalization, bias-field correction, spatial resampling, ComBat harmonization, protocol coordination, phantom scanning, and QA monitoring; and (3) Post-Harmonization, which focuses on data validation, assessment of preserved biological signal, and documentation of the harmonization process. This framework provides actionable methodological steps to support reproducible multicenter imaging research. QA: quality assurance; DICOM: Digital Imaging and Communications in Medicine; PET-CT: positron emission tomography-computed tomography.

Discussion

Although standardization efforts have been in place for a long time and may need to be further strengthened and expanded to better include radiomics, their ability to reduce variations in radiomic feature distributions across sites is still limited [33], and the preservation of fine anatomical detail and clinically relevant predictive information is essential in medical imaging, and the downstream impacts of harmonization must be evaluated with caution [34]. The main reason is the continuing diversity of scanner models, proprietary reconstruction algorithms, and post-processing tools used in different clinical centers [33]. In particular, several key challenges were identified in developing deep learning models using multi-site structural brain MRI datasets. These challenges can be grouped into four main categories: (1) difficulty in locating relevant literature, (2) limited access to suitable datasets, (3) a widespread lack of annotation in large datasets, and (4) the need to navigate the trade-off between data harmonization and domain adaptation strategies [35].

Conclusion

The harmonization of multicenter imaging data is indispensable for advancing epidemiologic research. Technical heterogeneity introduced by scanner manufacturers, acquisition protocols, reconstruction algorithms, and operator-dependent factors significantly compromises the reproducibility and generalizability of imaging biomarkers. By systematically addressing these challenges through pre-acquisition standardization, phantom-based calibration, post-acquisition normalization, and statistical or machine learning methods, researchers can substantially reduce site-related variability while preserving biologically meaningful signals. The integration of harmonization workflows into study design not only strengthens causal inference and statistical power but also facilitates collaboration across institutions and populations.

Looking ahead, future directions should emphasize the development of international standards, the incorporation of radiomics and deep learning into harmonization pipelines, and the adoption of federated learning frameworks that allow data sharing without compromising privacy. These efforts will expand the reach of imaging epidemiology, enabling robust and reproducible insights into population health.