多发性硬化患者自动病灶评估的前瞻性临床评价
前瞻性纳入2024年9月至12月常规随访MRI的多发性硬化患者112例,由4名神经放射科医生在有和无AI辅助下独立评估,评估AI对评估时间、工作流程及新发和增大病灶检出准确性的影响。AI辅助使平均评估时间减少27秒(p=0.317),放射科医生认为AI在87%的病例中有帮助,AI检出 new lesions 的阴性预测值为0.89,因假阳性倾向阳性预测值较低(0.35-0.65)。
要点
- AI辅助使平均评估时间减少27秒(p=0.317),差异未达统计学显著。
- 放射科医生认为AI在87%的病例中有帮助,11%的病例报告有困难。
- AI检出 new lesions 的阴性预测值为0.89,阳性预测值较低(0.35-0.65)。
核心结论
- AI工具在多发性硬化MRI随访中显示适度的、非显著的时间节省和较低的阳性预测值,限制了研究用途;高阴性预测值支持分诊潜力,放射科医生认为AI有助于病灶计数和检出小的新发病灶。
研究设计与样本
- 前瞻性临床验证研究,纳入2024年9月至12月常规随访MRI的多发性硬化患者112例,使用8种不同MRI扫描仪型号和不同协议采集,由4名神经放射科医生在有和无AI辅助(mdbrain v.4.11.0)下独立评估匿名化的当前和既往MRI检查,并完成效用问卷。
诊断标准与关键数值
- 以病例为单位计算检出 new and enlarging lesions 的敏感度、特异度和预测值;AI检出 new lesions 的阴性预测值为0.89(与无AI评估比较),阳性预测值为0.35-0.65。
临床应用与阅片要点
- 放射科医生认为AI工具在87%的病例中有帮助,在11%的病例中报告有困难;高阴性预测值支持分诊潜力,但低阳性预测值限制了研究用途。
局限性
- 时间节省为适度的、非显著的(平均减少27秒,p=0.317)。
- 阳性预测值较低(0.35-0.65),原文指出因假阳性倾向所致。
- 原文指出需要在缺乏明确 ground truth 的区域进行彻底的临床评估。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
PURPOSE: To perform a real-world clinical validation of a commercial AI tool for automatic MRI assessment in multiple sclerosis (MS) patients, evaluating its impact on assessment time, workflow, and accuracy in detecting new and enlarging lesions.
METHODS: We prospectively enrolled MS patients undergoing routine follow-up MRI from September-December 2024. Current and prior MRI examinations were anonymized and assessed independently by four neuroradiologists with and without AI assistance (mdbrain v.4.11.0). Assessment times were recorded, and radiologists completed utility questionnaires. Lesion quantification was compared between radiologist alone, radiologist with AI, and AI alone. Performance metrics including sensitivity, specificity, and predictive values were calculated case-level for detecting new and enlarging lesions.
RESULTS: The cohort included 112 MS patients scanned on 8 different MRI scanner models with varying protocols. Mean assessment time was reduced by 27 s when using AI versus without (p = 0.317). Radiologists found AI helpful in 87% of cases and reported difficulties in 11%. AI obtained negative predictive values of 0.89 for detecting new lesions when comparing to assessment without AI. Positive predictive values were low (0.35-0.65) due to false positive tendencies.
CONCLUSION: We prospectively validated an AI tool for MS MRI follow-up in a real-world setting. It showed modest, non-significant time savings and low positive predictive value, limiting research use. High negative predictive value supports triaging potential. Radiologists found the AI tool helpful for lesion counting and detecting small new lesions. Findings highlight the need for thorough clinical evaluation, especially in areas lacking definitive ground truth.
引用
- Hindsholm AM, Langkilde AR, Carlsen JF, et al. Prospective clinical evaluation of automatic lesion assessment in patients with multiple sclerosis. Neuroradiology. 2025. PMID: 41222674. doi:10.1007/s00234-025-03834-4
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。