Department of Neuromuscular Diseases, UCL Queen Square Institute of Neurology, London, UK.
UCL Centre for Medical Image Computing, London, UK.
J Neurol Neurosurg Psychiatry. 2024 May 14;95(6):500-503. doi: 10.1136/jnnp-2023-332454.
Lower limb muscle magnetic resonance imaging (MRI) obtained fat fraction (FF) can detect disease progression in patients with Charcot-Marie-Tooth disease 1A (CMT1A). However, analysis is time-consuming and requires manual segmentation of lower limb muscles. We aimed to assess the responsiveness, efficiency and accuracy of acquiring FF MRI using an artificial intelligence-enabled automated segmentation technique.
We recruited 20 CMT1A patients and 7 controls for assessment at baseline and 12 months. The three-point-Dixon fat water separation technique was used to determine thigh-level and calf-level muscle FF at a single slice using regions of interest defined using Musclesense, a trained artificial neural network for lower limb muscle image segmentation. A quality control (QC) check and correction of the automated segmentations was undertaken by a trained observer.
The QC check took on average 30 seconds per slice to complete. Using QC checked segmentations, the mean calf-level FF increased significantly in CMT1A patients from baseline over an average follow-up of 12.5 months (1.15%±1.77%, paired t-test p=0.016). Standardised response mean (SRM) in patients was 0.65. Without QC checks, the mean FF change between baseline and follow-up, at 1.15%±1.68% (paired t-test p=0.01), was almost identical to that seen in the corrected data, with a similar overall SRM at 0.69.
Using automated image segmentation for the first time in a longitudinal study in CMT, we have demonstrated that calf FF has similar responsiveness to previously published data, is efficient with minimal time needed for QC checks and is accurate with minimal corrections needed.
下肢肌肉磁共振成像（MRI）获得的脂肪分数（FF）可检测到 Charcot-Marie-Tooth 病 1A（CMT1A）患者的疾病进展。然而，分析既耗时又需要手动分割下肢肌肉。我们旨在评估使用人工智能辅助自动分割技术获取 FF MRI 的响应性、效率和准确性。
我们招募了 20 名 CMT1A 患者和 7 名对照者进行基线和 12 个月的评估。使用三点 Dixon 脂肪水分离技术，在单个切片上使用通过 Musclesense 定义的感兴趣区域来确定大腿水平和小腿水平的肌肉 FF，Musclesense 是一种用于下肢肌肉图像分割的经过训练的人工神经网络。通过经过培训的观察者进行质量控制（QC）检查和自动分割的校正。
QC 检查平均每片需要 30 秒才能完成。使用经过 QC 检查的分割，CMT1A 患者的小腿水平 FF 在平均 12.5 个月的随访中从基线显著增加（1.15%±1.77%，配对 t 检验 p=0.016）。患者的标准化反应均值（SRM）为 0.65。没有 QC 检查，在基线和随访之间，FF 的平均变化为 1.15%±1.68%（配对 t 检验 p=0.01），与校正后的数据几乎相同，整体 SRM 为 0.69。
在 CMT 的首次纵向研究中，我们首次使用自动图像分割，证明了小腿 FF 与之前发表的数据具有相似的响应性，具有高效性，仅需很少的 QC 检查时间，并且准确性高，仅需很少的校正。
