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.

Graphical Abstract