自动化深度学习辅助MRI早期检测放射性颞叶损伤:一项多中心回顾性分析
该多中心回顾性研究纳入4家医院2010年1月至2021年6月接受放疗和MRI随访的6483例鼻咽癌患者中的396例放射性颞叶损伤(RTLI)和3181例非RTLI患者,开发并验证了基于Unet++颞叶分割和改良DenseNet-121诊断的自动化工具RTLI-DM,多读者研究显示其辅助可显著提高4名放射科医生(2名培训期、2名经验丰富)的敏感度并减少漏诊,培训期医生特异度略降、假阳性略增,平均读片时间缩短29秒。
要点
- RTLI-DM辅助使培训期医生敏感度从60.5%升至83.5%、经验丰富医生从79.5%升至91.5%,漏诊均减少。
- 培训期医生辅助后特异度从99.0%降至93.0%、假阳性从0.01升至0.07(p < 0.05)。
- 辅助读片平均时间较未辅助缩短29秒,但原文指出临床应用时仍需验证。
核心结论
- 对于MRI上RTLI的早期检测,RTLI-DM辅助显著提高了放射科医生的诊断性能,同时缩短了读片时间。
研究设计与样本
- 多中心回顾性分析,从4家医院2010年1月至2021年6月接受放疗和MRI随访的6483例鼻咽癌患者中纳入396例RTLI和3181例非RTLI患者。
诊断标准与关键数值
- RTLI-DM由基于Unet++的双侧颞叶分割模型和基于改良DenseNet-121网络的诊断模型组成。
- 培训期医生未辅助敏感度60.5%(95% CI: 53.6-67.0%) vs 辅助后83.5%(95% CI: 77.7-88.0%)。
- 经验丰富医生未辅助敏感度79.5%(95% CI: 73.3-84.5%) vs 辅助后91.5%(95% CI: 86.7-94.7%)。
- 培训期医生未辅助特异度99.0%(95% CI: 96.2-100%) vs 辅助后93.0%(95% CI: 88.5-95.9%),辅助后假阳性从0.01升至0.07(p < 0.05)。
- 培训期医生未辅助假阴性0.395 vs 辅助后0.165;经验丰富医生未辅助假阴性0.205 vs 辅助后0.085。
临床应用与阅片要点
- RTLI-DM可作为诊断辅助工具用于MRI上RTLI的早期检测,尤其有助于发现细微病灶。
- 原文指出RTLI-DM在临床应用中仍需验证。
局限性
- 原文指出RTLI-DM在临床应用中仍需验证。
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
OBJECTIVES: To evaluate the benefits of an automated deep learning-based tool (RTLI-DM) for early detection of radiation-induced temporal lobe injury (RTLI) on MRI.
MATERIALS AND METHODS: A total of 396 RTLI and 3181 non-RTLI patients were retrospectively included from a cohort of 6483 NPC patients who underwent radiotherapy and MRI follow-ups at four hospitals from January 2010 to June 2021. We developed and validated an automated RTLI-DM using the initial abnormal MRI of RTLI patients and the final MRI of non-RTLI patients. The RTLI-DM consists of a bilateral temporal lobe segmentation model based on Unet++ and a diagnostic model based on a modified DenseNet-121 network. The multi-reader study was conducted to compare the performance of four radiologists (two trainees, and two experienced readers) in interpreting images without and with RTLI-DM assistance. Sensitivity, specificity, false positives (FPs), and false negatives (FNs) were compared between two reading sessions.
RESULTS: Assisted reading with RTLI-DM significantly improve the sensitivity (60.5% [95% CI: 53.6-67.0%] vs 83.5% [95% CI: 77.7-88.0%]; 79.5% [95% CI: 73.3-84.5%] vs 91.5% [95% CI: 86.7-94.7%]) and FNs (0.395 vs 0.165; 0.205 vs 0.085) for both trainees and experienced readers, although there is a slight decrease in specificity (99.0% [95% CI: 96.2-100%] vs 93.0% [95% CI: 88.5-95.9%]) and increase in FPs (0.01 vs 0.07) for trainees (p < 0.05). The averaged time for assisted reading is reduced by 29 s, compared with unassisted reading.
CONCLUSION: For early detection of RTLI on MRI, the RTLI-DM assistant significantly improved the diagnostic performance of radiologists while reducing the reading time.
KEY POINTS: Question RTLI is usually asymptomatic in its initial stage, which contributes to the challenges of early diagnosis. Findings The effectiveness of the RTLI detection model (RTLI-DM) as a diagnostic aid is evident, particularly in enhancing the detection of subtle lesions. Clinical relevance The RTLI-DM can significantly improve radiologists' diagnostic efficiency and accuracy in the early detection of RTLI in clinical workflows, but still requires validation when in clinical applications.
引用
- Yang F, Hu R, Hu J, et al. Automated deep learning-assisted early detection of radiation-induced temporal lobe injury on MRI: a multicenter retrospective analysis. Eur Radiol. 2025. PMID: 40050455. doi:10.1007/s00330-025-11470-y
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。