Purpose:
Differentiating active from non-active multiple sclerosis (MS) lesions is critical for disease management but often relies on gadolinium-enhanced magnetic resonance imaging (MRI), raising concerns about retention risks and costs. This study introduces a contrast-free, multi-sequence MRI approach using radiomics and machine learning to classify MS lesion activity.

Material and methods:
A total of 187 lesions from 31 MS patients (mean age 42.5 ± 11.3 years; 64.5% female) at Amin Hospital (November 2024 – February 2025) were retrospectively analysed using a 1.5 T MRI scanner. Five sequences – T1-weighted (T1W), T2-weighted (T2W), fluid-attenuated inversion recovery (FLAIR), diffusion-weighted imaging (DWI), and susceptibility-weighted imaging (SWI) – were processed to extract 8905 radiomic features, refined to 127 via correlation and recursive feature elimination. XGBoost classified lesions as active or non-active, validated on an internal test set (n = 28 lesions), with performance assessed by area under the receiver operating characteristic curve (AUC-ROC).

Results:
The XGBoost model achieved an AUC-ROC of 0.87 (95% CI: 0.82-0.92), sensitivity of 0.85, and specificity of 0.83, outperforming other classifiers (SVM AUC 0.84). FLAIR (35.4%) and T2W (28.3%) dominated feature contributions, with SWI (12.6%) enhancing accuracy (AUC dropped to 0.84 without SWI). Noise simulation (Gaussian σ = 0.1) confirmed robustness (AUC = 0.86).

Conclusions:
This integration of SWI with conventional sequences in a unified radiomic model offers a promising contrast-free alternative for MS lesion classification, achieving promising accuracy comparable to radiologist performance on an internal test set (n = 28 lesions), pending external validation. External validation is needed to confirm the genera­lisability, but this approach could reduce gadolinium reliance in clinical practice.