# Stability of dynamic radiomics features in cardiac MRI under noise

## TL;DR

This study examines how stable radiomic features in cardiac MRI are when noise is introduced, finding that some dynamic features remain stable despite noise.

## Contribution

The paper introduces a novel evaluation of radiomic feature stability under noise in dynamic cardiac MRI data.

## Key findings

- Feature stability ranged from near 0 to over 20, with most features below 2.5 in mean pairwise MAE.
- GLSZM features showed lower stability compared to first-order features.
- Some features were sensitive to noise levels but stable across noise realizations at the same level.

## Abstract

Radiomic studies on cardiac MRI mainly focus on images from distinct time points rather than considering the system’s dynamic nature. Recent studies have shown that radiomic features exhibit considerable variation across the cardiac cycle and that dynamic features can improve classification accuracy in downstream tasks. However, it is unclear whether the dynamic temporal evolution of radiomic features is sufficiently stable in the presence of noise. In this work, we evaluate the stability of radiomic feature curves of cine CMR images under noise.

We extracted 910 radiomic features from all time points of cine CMR images of 115 subjects from three cohorts with various levels of artificially added noise. The stability of feature curves is evaluated based on pairwise normalized mean absolute errors, and features are ranked by their stability. Feature stability, measured by mean pairwise MAE, ranged from near 0 to over 20, with most features showing values below 2.5. Stability rankings showed moderate consistency across subjects (median Spearman correlation coefficient of 0.58). Features from the grey level size zone matrix (GLSZM) category demonstrated lower stability compared to first-order features. Some features exhibited high sensitivity to noise level but remained stable across different noise realizations at the same level.

Some radiomic feature curves remain stable under noise while showing variability over the cardiac cycle. These features are promising candidates for improving models using dynamic rather than static feature values.

Graphical Abstract