Evaluating Multi-Channel Input of Sagittal T2-weighted and STIR MRI for Automated Segmentation of Spinal Cord MS Lesions

Identifying multiple sclerosis lesions in spinal cord MRI is important but challenging, and automated methods have been proposed to reduce inter-rater variability. Although clinicians typically use multiple MR sequences to identify lesions for a given patient, most automated methods use only a single sequence as input. We investigated combining sagittal T2-weighted and STIR sequences using an early fusion U-Net for lesion segmentation, assessing the impact of resampling techniques and U-Net architectures (3D vs. hybrid 2D/3D). Results showed no overall improvement from adding STIR (p=0.17), though this varied among subjects. Furthermore, T2-only models trained on larger datasets outperformed T2+STIR models trained on subsets, indicating that a larger single-sequence dataset is preferable to a smaller multi-sequence one.