Multiple sclerosis and related disorders · 2026
Automatic multiple sclerosis lesion segmentation in the spinal cord using 3 T and 7 T MP2RAGE images.
СтатьяМногоцентровое исследованиеНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Laines-Medina Nilser, Mchinda Samira, Testud Benoit, Demortière Sarah, Pravatà Emanuele, Nair Govind, Reich Daniel S, Granziera Cristina, Tsagkas Charidimos, Callot Virginie, Cohen-Adad Julien
Аннотация
PURPOSE: To develop a deep learning-based fully automated framework for segmenting multiple sclerosis (MS) lesions in the spinal cord (SC) using images derived from MP2RAGE acquisitions.
MATERIALS AND METHODS: This retrospective multicenter study included 472 MRI volumes from 422 subjects (mean age 52±12.5 years), comprising 402 patients with MS and 20 healthy controls, acquired at three centers using 3 T and 7 T MRI systems. MP2RAGE-UNI images were preprocessed and analyzed using the Spinal Cord Toolbox (SCT). The multicenter 3 T dataset was divided into training (60%), validation (20%), and testing (20%) subsets, while the 7 T dataset was reserved for testing. The proposed method (UNIseg), based on the nnU-Net v2 architecture, was trained on the multicenter dataset and compared with the current benchmark (seg_ms_lesion_mp2rage, available in SCT v6.3 and earlier versions), which was trained on single-center MP2RAGE-UNI data. Segmentation performance was evaluated using Dice coefficient and lesion-wise metrics by comparing both UNIseg and the benchmark against reference standard. Statistical comparisons were performed using the Wilcoxon signed-rank test.
RESULTS: On the 3 T dataset, UNIseg achieved a mean Dice score of 0.67, significantly outperforming the benchmark (p ≤ 0.01). Lesion-wise sensitivity and precision were 0.81 and 0.92, respectively, with fewer false-positive segmentations. On the 7 T dataset, UNIseg also demonstrated significantly higher performance (p ≤ 0.001), although segmentation performance was lower than that observed at 3 T CONCLUSION: UNIseg provides a robust and efficient approach for segmentation of lesions, improving accuracy compared with existing methods, especially for 3 T data. The trained model and code are publicly available in SCT v6.4 and later.
MATERIALS AND METHODS: This retrospective multicenter study included 472 MRI volumes from 422 subjects (mean age 52±12.5 years), comprising 402 patients with MS and 20 healthy controls, acquired at three centers using 3 T and 7 T MRI systems. MP2RAGE-UNI images were preprocessed and analyzed using the Spinal Cord Toolbox (SCT). The multicenter 3 T dataset was divided into training (60%), validation (20%), and testing (20%) subsets, while the 7 T dataset was reserved for testing. The proposed method (UNIseg), based on the nnU-Net v2 architecture, was trained on the multicenter dataset and compared with the current benchmark (seg_ms_lesion_mp2rage, available in SCT v6.3 and earlier versions), which was trained on single-center MP2RAGE-UNI data. Segmentation performance was evaluated using Dice coefficient and lesion-wise metrics by comparing both UNIseg and the benchmark against reference standard. Statistical comparisons were performed using the Wilcoxon signed-rank test.
RESULTS: On the 3 T dataset, UNIseg achieved a mean Dice score of 0.67, significantly outperforming the benchmark (p ≤ 0.01). Lesion-wise sensitivity and precision were 0.81 and 0.92, respectively, with fewer false-positive segmentations. On the 7 T dataset, UNIseg also demonstrated significantly higher performance (p ≤ 0.001), although segmentation performance was lower than that observed at 3 T CONCLUSION: UNIseg provides a robust and efficient approach for segmentation of lesions, improving accuracy compared with existing methods, especially for 3 T data. The trained model and code are publicly available in SCT v6.4 and later.