Reproducibility and discriminative power of MRI liver radiomics: Impact of deep learning-based image reconstruction

While Deep Learning-based Image Reconstruction (DLIR) is increasingly used to improve MRI quality, its influence on radiomic stability remains underexplored. This study aimed to assess the impact of DLIR on the reproducibility and discriminative power of radiomic features in liver MRI.

In this retrospective study, 95 patients with malignant liver lesions underwent 1.5 T MRI (T2, T2FS, and Diffusion). Images were reconstructed using both traditional (non-DLIR) and DLIR (AIR Recon DL) methods. Three regions of interest (ROI) - tumor, peritumoral, and healthy liver - were segmented by two readers. Intra- and inter-observer reproducibility were evaluated using ICC. Discriminative power was assessed via Random Forest classifiers and Wilcoxon tests to distinguish tissue types.

Generalized Linear Mixed Models showed no significant difference in the number of reproducible features between DLIR and non-DLIR for intra-observer (Odds Ratio (OR): 0.860, 95%CI [0.653; 1.135]) or inter-observer (OR: 1.098, 95%CI [0.673; 1.791]) analyses. Inter-observer reproducibility remained low across all reconstructions, particularly for intensity features. Classifier performance for tissue differentiation remained comparable between reconstructions. A slight decrease (2-7%) in the percentage of discriminative features was observed with DLIR, suggesting a mild smoothing effect on subtle textures.

DLIR does not significantly improve or degrade the reproducibility of radiomics in liver MRI. However, DLIR preserves the essential signatures required for tissue discrimination. These results support the clinical integration of DLIR in radiomic workflows, confirming that it can be used without compromising the reliability of quantitative imaging biomarkers.