文献速递腹部技术与安全CTMRUS

面向定量MASLD生物标志物映射的物理感知影像人工智能:超声、CT和MRI深度学习和放射组学的系统综述

Physics-aware imaging AI for quantitative MASLD biomarker mapping: a systematic review of deep learning and radiomics across ultrasound, CT, and MRI.
PubMed 文献 · Abdom Radiol (NY) · 2025-11-26

该系统综述纳入2015至2025年间33项研究,评估物理感知AI在MASLD/MASH定量生物标志物映射中的应用。MRI在脂肪/铁定量方面领先,CT在机会性脂肪评估方面可扩展性强,超声方法通用性正在提高。部署就绪度最高的是CT衰减和MRI PDFF/R2*流程。

要点

  • MRI使用PDFF和R2*进行生物物理基础的脂肪/铁定量,并采用自动全肝分割。
  • CT通过全自动体积衰减和双能虚拟非对比进行机会性脂肪评估,但性能依赖于期相。
  • 超声分为物理知情定量超声(衰减/背散射)与B型深度学习脂肪分级,域适应技术正在兴起。

核心结论

  • MASLD/MASH的影像AI正趋向物理一致、自动化的定量映射。
  • 部署就绪度最高的是CT衰减和MRI PDFF/R2*流程。
  • 未来需要前瞻性、多中心验证,整合物理感知设计、严格混杂控制和标准化报告。

研究设计与样本

  • PRISMA 2020指导的系统综述,检索PubMed、Scopus和Web of Science,时间范围为2015至2025年。
  • 共纳入33项研究,纳入标准为应用AI进行自动分割、定量脂肪/铁/纤维化映射或MASLD/MASH分期的影像研究。
  • 提取并叙述性合成技术方法、物理感知设计、参考标准、性能和部署就绪度等数据。

关键影像表现/征象

  • MRI采用混杂校正序列和自动全肝分割进行PDFF和R2*定量,在生物物理基础的脂肪/铁定量方面领先。
  • CT通过全自动体积衰减和双能虚拟非对比进行机会性脂肪评估,可扩展性强,但性能依赖于期相。
  • 超声分为物理知情定量超声(衰减/背散射)与MRI-PDFF相关,以及B型深度学习用于脂肪分级,并出现域适应技术。

诊断标准与关键数值

  • CT和MRI自动化的证据强度和外部验证最为稳健,超声方法的通用性正在提高。
  • 关键差距包括标准化纵向流程、多厂商协调和MASLD特异性纤维化验证。

临床应用与阅片要点

  • CT衰减和MRI PDFF/R2*流程的部署就绪度最高,可优先用于MASLD/MASH的定量生物标志物映射。
  • 超声定量方法(衰减/背散射)与MRI-PDFF相关,B型深度学习可用于脂肪分级,但通用性仍在发展中。
  • 应用时需注意CT性能依赖于期相,并关注标准化纵向流程和多厂商协调等尚未解决的差距。

局限性

  • 关键差距包括标准化纵向流程、多厂商协调和MASLD特异性纤维化验证。
  • 未来工作需要前瞻性、多中心验证,整合物理感知设计、严格混杂控制和标准化报告。

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

OBJECTIVE: This systematic review critically appraises the current landscape of physics-aware artificial intelligence (AI) in medical imaging for quantitative biomarker mapping in Metabolic dysfunction-associated steatotic liver disease (MASLD) and its progressive form, MASH. It focuses on deep learning and radiomics applications across ultrasound, CT, and MRI.

METHODS: A PRISMA 2020-guided systematic review was conducted, searching PubMed, Scopus, and Web of Science from 2015 to 2025. Studies applying AI to imaging for automated segmentation, quantitative steatosis/iron/fibrosis mapping, or staging in MASLD/MASH were included. Data on technical approaches, physics-aware design, reference standards, performance, and deployment readiness were extracted and synthesized narratively.

RESULTS: Of 842 identified records, 33 studies were included. MRI leads in biophysically-grounded fat/iron quantification (PDFF, R2*) using confounder-corrected sequences and automated whole-liver segmentation. CT excels in scalable, opportunistic steatosis assessment via fully automated volumetric attenuation and dual-energy virtual non-contrast, though performance is phase-dependent. Ultrasound bifurcates into physics-informed quantitative ultrasound (attenuation/backscatter) correlating with MRI-PDFF and B-mode deep learning for steatosis grading, with emerging domain-adaptation techniques. Evidence strength and external validation are most robust for CT and MRI automation, while ultrasound methods are advancing in generalizability. Key gaps include standardized longitudinal pipelines, multi-vendor harmonization, and MASLD-specific fibrosis validation.

CONCLUSION: Imaging AI for MASLD/MASH is converging on physics-consistent, automated quantitative mapping. Deployment readiness is highest for CT attenuation and MRI PDFF/R2* pipelines. Future work requires prospective, multicenter validation integrating physics-aware design, rigorous confounder control, and standardized reporting to enable reliable clinical integration.

来源:Abdom Radiol (NY),查看原文。版权归原作者所有。

引用

  1. Maghsoudi H, Khonche A, Gereami R, et al. Physics-aware imaging AI for quantitative MASLD biomarker mapping: a systematic review of deep learning and radiomics across ultrasound, CT, and MRI. Abdom Radiol (NY). 2025. PMID: 41291295. doi:10.1007/s00261-025-05317-9

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。