Automated segmentation of multiple sclerosis lesions using 7 Tesla MRI multi-contrast data

Accurate detection and quantification of lesions in multiple sclerosis remain challenging, which hampers precise monitoring and may impact treatment decisions. In clinical practice, lesion segmentation is mostly performed manually, making the process labor-intensive, time-consuming, and prone to subjective errors. This study proposes a deep learning-based method of multiple sclerosis lesions segmentation using a multi-contrast 7T MRI database. The approach was evaluated on a 7T MRI database, using single and multi-contrast inputs and comparing the results with both conventional algorithms and deep learning-based methods. When using a single contrast, the highest DSC was achieved with MP2RAGE (DSC = 0.66), while FLAIR showed a lower DSC = 0.48. The combination of MP2RAGE and FLAIR provided the best performance, with DSC of 0.70. The proposed method outperformed existing traditional and deep learning techniques, demonstrating the potential of high-field multi-contrast MRI for automated MS lesion segmentation. These results highlight the importance of high-quality ground truth segmentation, due to potential inconsistencies, as well as the value of an expanded dataset for further improving diagnostic accuracy in clinical practice.