Impact of different pre-processing methods combined with combat harmonization on the reproducibility of radiomics...

To explore the effect of different pre-processing methods combined with ComBat harmonization on improving the stability of MRI radiomic features in healthy human brains.

T1FLAIR and T2W scans from 416 healthy subjects across four MRI systems underwent seven pipelines (Original, N4, Z-Score, DTN, Nyul, WhiteStripe, RAVEL). SynthSeg segmented white/gray matter for 93 IBSI radiomic features. Stability was quantified via covariate-adjusted standardized scanner effect (d = β_scanner/σ_res, |d|< 0.2) with 95% bootstrap CIs, supplemented by PCA and LOOCV AUC. Friedman and post-hoc McNemar tests with Bonferroni correction compared pipelines. FDR-corrected partial correlations assessed retention of significant age and sex associations after ComBat.

Raw data exhibited extremely low stability with substantial scanner-induced effects, while scanner classification AUC was close to 1.0. ComBat harmonization significantly improved stability across all pipelines and reduced AUC to near 0.5. Within this dataset, DTN_ComBat demonstrated consistently high stability; pairwise statistical comparisons indicated it showed more favourable stability performance relative to most other pipelines for the majority of tested contrasts, with stable proportions exceeding 90% across all sequences and tissues, and retained 75.0-100.0% of statistically significant age and sex-related associations.

Unprocessed brain MRI radiomic features exhibit poor cross‑scanner stability. Within this dataset, ComBat effectively reduces scanner‑related feature‑distribution differences, while preserving most statistically‑significant age and sex‑related associations. Among tested pre-processing pipelines, DTN_ComBat showed comparatively favourable performance under the present healthy‑brain MRI scanner‑protocol configurations.