Integrating Radiomics with Deep Learning Enhances Multiple Sclerosis Lesion Delineation
Authors
Abstract
Background: Accurate lesion segmentation is critical for multiple sclerosis (MS) diagnosis, yet current deep learning approaches face robustness challenges. Aim: This study improves MS lesion segmentation by combining data fusion and deep learning techniques. Materials and Methods: We suggested novel radiomic features (concentration rate and R\'enyi entropy) to characterize different MS lesion types and fused these with raw imaging data. The study integrated radiomic features with imaging data through a ResNeXt-UNet architecture and attention-augmented U-Net architecture. Our approach was evaluated on scans from 46 patients (1102 slices), comparing performance before and after data fusion. Results: The radiomics-enhanced ResNeXt-UNet demonstrated high segmentation accuracy, achieving significant improvements in precision and sensitivity over the MRI-only baseline and a Dice score of 0.774$\pm$0.05; p<0.001 according to Bonferroni-adjusted Wilcoxon signed-rank tests. The radiomics-enhanced attention-augmented U-Net model showed a greater model stability evidenced by reduced performance variability (SDD = 0.18 $\pm$ 0.09 vs. 0.21 $\pm$ 0.06; p=0.03) and smoother validation curves with radiomics integration. Conclusion: These results validate our hypothesis that fusing radiomics with raw imaging data boosts segmentation performance and stability in state-of-the-art models.