Abstract

Accurate detection and delineation of white matter lesions remain critical for diagnosing and monitoring multiple sclerosis (MS), a chronic autoimmune disease that affects nearly three million people worldwide. Traditional assessments based on MRI, particularly fluid-attenuated inversion recovery (FLAIR) sequences, rely heavily on qualitative interpretation and are prone to observer variability. While deep learning (DL) models have substantially advanced automated lesion segmentation, these approaches often lack robustness and exhibit unstable training behaviors, undermining their reliability in clinical practice. Radiomics, which involves extracting high-dimensional quantitative features that characterize lesion texture and intensity distributions, has shown potential to enrich neuroimaging analyses. However, the integration of explicit radiomic descriptors into DL segmentation pipelines for MS lesion detection has not been systematically explored. This study aimed to determine whether fusing two quantitative radiomic features—Concentration Rate (CR), capturing local hyperintensity clustering, and Rényi Entropy (RE), quantifying textural disorder—alongside raw MRI data could enhance both segmentation performance and training stability in advanced DL models. To address this, we retrospectively analyzed FLAIR MRI scans from 46 patients diagnosed with MS at Tawam Hospital in Al Ain, comprising a total of 1102 annotated slices. CR was computed to highlight local intensity aggregations by summing high but not extreme voxel values within a scanning window, while RE measured the complexity of intensity patterns using grey-level co-occurrence matrices, emphasizing subtle heterogeneities that often characterize neuroinflammatory lesions. These radiomic features were integrated with the raw FLAIR slices and fed into two state-of-the-art DL segmentation architectures: a ResNeXt-UNet, which combines a ResNeXt-50 encoder for powerful feature extraction with a classic U-Net decoder, and an attention-augmented U-Net that leverages squeeze-and-attention blocks to refine spatial focus. Each model was trained under two conditions—first using only raw MRI data, and then using MRI data fused with CR and RE features. We employed six-fold cross-validation to rigorously evaluate segmentation performance, measuring Dice similarity coefficient, precision, and sensitivity, while also calculating the standard deviation of derivatives (SDD) of validation scores across epochs to quantify the stability of the training process. Statistical significance was assessed using Wilcoxon signed-rank tests with Bonferroni adjustments for multiple comparisons. Our findings demonstrated that incorporating CR and RE significantly improved both the accuracy and robustness of MS lesion segmentation. In the ResNeXt-UNet pipeline, the Dice score increased from 0.680±0.041 with MRI alone to 0.706±0.046 with the addition of radiomic features (p<0.001), accompanied by gains in precision and sensitivity. The attention-augmented U-Net exhibited similar improvements, with Dice rising from 0.707±0.054 to 0.719±0.063 (p<0.001), alongside higher precision and sensitivity metrics. Notably, the integration of radiomic features also reduced the SDD from 0.21±0.06 to 0.18±0.09 (p=0.03) in the attention-augmented model, indicating a smoother and more stable learning curve that mitigates overfitting risks. Visual inspection of validation trajectories confirmed these quantitative results, with radiomics-enhanced models showing fewer fluctuations and more consistent convergence. In conclusion, this study provides compelling evidence that explicitly fusing carefully selected radiomic features with raw MRI data enhances both the segmentation accuracy and training stability of advanced DL models for MS lesion delineation. By enriching the input space with descriptors that capture lesion-specific intensity clustering and textural complexity, these models become better equipped to identify subtle pathological variations that are often missed by purely intensity-based approaches. Such radiomics-informed AI frameworks hold promise for improving the objectivity and reproducibility of MS lesion assessments, ultimately supporting more precise application of diagnostic criteria and longitudinal monitoring of disease progression. Future research should aim to validate these findings across larger multicenter datasets and explore additional data fusion strategies incorporating multimodal imaging or clinical variables to further advance precision neuroimaging in demyelinating disorders.