**Background:** UK Biobank (UKBB) and the German National Cohort (NAKO) are two of the largest population-scale cohort studies, each collecting MRI data alongside extensive health-related information. UKBB enrolls adults aged 50–80 years, while NAKO enrolls participants aged 20–70 years, limiting the generalizability of each study individually. Merging data across studies could overcome these limitations and increase statistical power, but differences in scanner types, field strengths (1.5 T in UKBB vs. 3 T in NAKO), imaging protocols, and voxel sizes introduce distribution shifts that must be addressed. This study aimed to demonstrate how imaging data from UKBB and NAKO can be jointly analyzed after harmonization and to derive comprehensive quantitative image-based organ phenotypes across the general adult population.
**Methods:** T1-weighted Dixon MRI data from 10,000 participants per study were processed. After excluding scans with artifacts or erroneous segmentations, 17,996 participants (8791 UKBB, 9205 NAKO) were included. Deep learning-based organ segmentation (nnUNet) was used to segment liver, spleen, left and right kidneys, and pancreas. Kidney hilum adipose tissue (AT) was separated from parenchyma using a fat-signal threshold of 0.5. Proton density fat fractions (PDFF) of liver and pancreas were estimated from fat and water signal intensities. Nine image-derived features were extracted per participant. The data-generating process was analyzed using prior knowledge combined with nonparametric nonlinear conditional independence testing (Invariant Environment Prediction with random forests, 100 trees, 5-fold cross-validation, significance threshold 0.01 with Holm-Bonferroni correction). Data harmonization was performed using ComBat, with data source (UKBB vs. NAKO) as the batch variable and age, sex, height, and weight as covariates to be preserved. UKBB was used as the reference dataset. After harmonization, multilinear quantile regression (with a quadratic age term) was used to model the relationship between demographic variables and image-derived features.
**Results:** UKBB participants were older on average (63.0 ± 7.5 years) than NAKO participants (51.8 ± 11.4 years). Sex was balanced in both cohorts (49.9% female in UKBB, 49.1% female in NAKO). Causal discovery revealed direct effects of sex (P < 0.0001), height (P < 0.0001), weight (P < 0.0001), and data source (P < 0.0001) on image-derived features, but no direct effect of data source on weight (P = 0.95) or height (P = 0.99) beyond age mediation. ComBat harmonization improved alignment of marginal feature densities between studies, particularly for pancreas volume and liver PDFF. For some features (e.g., pancreas PDFF, right kidney AT volume), harmonization increased apparent differences, which the authors attribute to preservation and enhancement of age-related effects. After harmonization, cross-study analysis showed a marked nonlinear decrease in organ volumes with age (steepest between 40–80 years) and a substantial increase in kidney hilum AT volumes with age. Liver PDFF and pancreas PDFF both increased nonlinearly with age, with the increase more pronounced for pancreas. Quantile regression models (Table 2) showed that only the quadratic age term and body weight had nonzero coefficients; height and linear age terms were not retained. For example, in females, liver PDFF (intercept −5.80E-03%) had coefficients of 2.80E-06 %/y² for age² and 2.40E-04 %/kg for weight; in males, liver PDFF (intercept −1.50E-02%) had coefficients of 3.00E-06 %/y² and 3.90E-04 %/kg. Liver volume in females (intercept 660 mL) changed by −2.10E-02 mL/y² and 11.0 mL/kg; in males (intercept 670 mL) by −4.00E-02 mL/y² and 13.0 mL/kg. A supplemental analysis showed that classification accuracy for predicting data source from image features markedly decreased after harmonization, indicating successful bias reduction.
**Clinical Implications:** This study provides a blueprint for cross-study harmonization of large-scale MRI data, demonstrating that ComBat combined with causal discovery can effectively correct for acquisition-related biases while preserving biological variation. The resulting quantitative models of abdominal organ volumes and fat distribution across a 20–80 year age range offer potential reference values for clinical diagnostics, such as defining normative ranges for liver fat fraction or kidney size. The ability to pool data from UKBB and NAKO increases statistical power and generalizability, enabling future studies linking genetic, environmental, and behavioral risk factors to MRI-derived phenotypes. Limitations include the reliance on T1-weighted Dixon images rather than dedicated multi-echo sequences for PDFF estimation, the absence of external calibration for the choice of reference dataset (UKBB), and the potential for unobserved confounders (e.g., ethnicity, lifestyle) to contribute to residual bias.