Normalization Strategies in Multi-Center Radiomics Abdominal MRI: Systematic Review and Meta-Analyses

Impact of MRI radiomic feature normalization for prognostic modelling in uterine endometrial and cervical cancers

Widespread clinical use of MRI radiomic tumor profiling for prognostication and treatment planning in cancers faces major obstacles due to limitations in standardization of radiomic features. The purpose of the current work was to assess the impact of different MRI scanning- and normalization protocols for the statistical analyses of tumor radiomic data in two patient cohorts with uterine endometrial-(EC) (n = 136) and cervical (CC) (n = 132) cancer. 1.5 T and 3 T, T1-weighted MRI 2 min post-contrast injection, T2-weighted turbo spin echo imaging, and diffusion-weighted imaging were acquired. Radiomic features were extracted from within manually segmented tumors in 3D and normalized either using z-score normalization or a linear regression model (LRM) accounting for linear dependencies with MRI acquisition parameters. Patients were clustered into two groups based on radiomic profile. Impact of MRI scanning parameters on cluster composition and prognostication were analyzed using Kruskal-Wallis tests, Kaplan-Meier plots, log-rank test, random survival forests and LASSO Cox regression with time-dependent area under curve (tdAUC) (α = 0.05). A large proportion of the radiomic features was statistically associated with MRI scanning protocol in both cohorts (EC: 162/385 [42%]; CC: 180/292 [62%]). A substantial number of EC (49/136 [36%]) and CC (50/132 [38%]) patients changed cluster when clustering was performed after z-score-versus LRM normalization. Prognostic modeling based on cluster groups yielded similar outputs for the two normalization methods in the EC/CC cohorts (log-rank test; z-score: p = 0.02/0.33; LRM: p = 0.01/0.45). Mean tdAUC for prognostic modeling of disease-specific survival (DSS) by the radiomic features in EC/CC was similar for the two normalization methods (random survival forests; z-score: mean tdAUC = 0.77/0.78; LRM: mean tdAUC = 0.80/0.75; LASSO Cox; z-score: mean tdAUC = 0.64/0.76; LRM: mean tdAUC = 0.76/0.75). Severe biases in tumor radiomics data due to MRI scanning parameters exist. Z-score normalization does not eliminate these biases, whereas LRM normalization effectively does. Still, radiomic cluster groups after z-score- and LRM normalization were similarly associated with DSS in EC and CC patients.

The effect of feature normalization methods in radiomics

OBJECTIVES

In radiomics, different feature normalization methods, such as z-Score or Min-Max, are currently utilized, but their specific impact on the model is unclear. We aimed to measure their effect on the predictive performance and the feature selection.

METHODS

We employed fifteen publicly available radiomics datasets to compare seven normalization methods. Using four feature selection and classifier methods, we used cross-validation to measure the area under the curve (AUC) of the resulting models, the agreement of selected features, and the model calibration. In addition, we assessed whether normalization before cross-validation introduces bias.

RESULTS

On average, the difference between the normalization methods was relatively small, with a gain of at most + 0.012 in AUC when comparing the z-Score (mean AUC: 0.707 ± 0.102) to no normalization (mean AUC: 0.719 ± 0.107). However, on some datasets, the difference reached + 0.051. The z-Score performed best, while the tanh transformation showed the worst performance and even decreased the overall predictive performance. While quantile transformation performed, on average, slightly worse than the z-Score, it outperformed all other methods on one out of three datasets. The agreement between the features selected by different normalization methods was only mild, reaching at most 62%. Applying the normalization before cross-validation did not introduce significant bias.

CONCLUSION

The choice of the feature normalization method influenced the predictive performance but depended strongly on the dataset. It strongly impacted the set of selected features.

CRITICAL RELEVANCE STATEMENT

Feature normalization plays a crucial role in the preprocessing and influences the predictive performance and the selected features, complicating feature interpretation.

KEY POINTS

• The impact of feature normalization methods on radiomic models was measured. • Normalization methods performed similarly on average, but differed more strongly on some datasets. • Different methods led to different sets of selected features, impeding feature interpretation. • Model calibration was not largely affected by the normalization method.