Deep learning MRI reconstruction significantly alters radiomic features and affects their reproducibility and repeatability

Deep learning (DL)-based reconstruction has been increasingly adopted in magnetic resonance imaging (MRI) to improve image quality; however, its influence on radiomic feature stability remains unclear. This exploratory study evaluated the impact of DL reconstruction on radiomic feature values, intraobserver reproducibility, and test–retest repeatability in normal canine brain MRI. T1-weighted, T2-weighted, and FLAIR sequences were acquired twice in nine healthy dogs, and radiomic features were extracted from six anatomical regions using conventional and DL-reconstructed images. Wilcoxon signed-rank tests, intraclass correlation coefficients (ICCs), and coefficients of variation (CVs) were used to assess value changes, reproducibility, and repeatability, respectively. DL reconstruction significantly altered most radiomic features across sequences and regions, suggesting that DL-reconstructed and non-DL-reconstructed images should not be assumed to be interchangeable for quantitative analysis. DL-reconstructed images generally showed higher intraobserver ICCs for multiple features and regions. In contrast, DL-reconstructed images also tended to show higher CVs across sequences, suggesting lower test–retest repeatability. These findings suggest that DL reconstruction may influence radiomic feature distributions and their stability, underscoring the need for careful validation and harmonization before incorporating DL-reconstructed MRI into radiomics workflows or model development.