文献速递神经CT

利用CT扫描深度学习预测脑挫伤水肿患者严重程度

Severity prediction in patients with oedema in cerebral contusion using deep learning from computed tomography scans.
PubMed 文献 · Clin Radiol · 2025-10-11

本研究开发结合深度学习的模型评估脑挫伤水肿严重程度。深度学习联合Tada公式与手动勾画量化141例创伤性颅内出血体积,另将脑挫伤水肿患者按21天格拉斯哥结局量表分为高危组56例和低危组66例,构建整合血肿与水肿的水肿指数(EI)。水肿体积(EV)预测水肿严重程度的AUC为0.840,高于Marshall CT分类的0.664、A5(+1) CT分类的0.692和EI的0.731(均P<.05);EI联合Marshall CT分类较单用Marshall CT分类AUC增加0.081(P=.021),EI联合A5(+1) CT分类较单用A5(+1) CT分类AUC增加0.075(P=.013)。结论认为,整合血肿与水肿的深度学习联合模型可用于预测脑挫伤水肿严重程度。

要点

  • 深度学习体积估测与手动勾画相比体积百分比误差为3.59%。
  • EV预测水肿严重程度AUC 0.840,高于Marshall CT分类0.664、A5(+1) CT分类0.692和EI 0.731(均P<.05)。
  • EI联合Marshall CT分类较单用Marshall CT分类AUC增加0.081(P=.021)。

核心结论

  • 整合血肿与水肿的深度学习联合模型可用于预测脑挫伤水肿的严重程度。

研究设计与样本

  • 研究论文;采用深度学习联合Tada公式与手动勾画量化141例创伤性颅内出血体积;另在脑挫伤水肿队列中按21天格拉斯哥结局量表分为高危组56例和低危组66例,构建整合血肿与水肿的水肿指数(EI),并与Marshall CT分类、A5(+1) CT分类、水肿体积(EV)比较评估水肿严重程度。

关键影像表现/征象

  • 深度学习体积估测与手动勾画相比体积百分比误差为3.59%。

诊断标准与关键数值

  • EV预测水肿严重程度的AUC为0.840,高于Marshall CT分类的0.664(P<.05)。
  • EV的AUC 0.840高于A5(+1) CT分类的0.692(P<.05)。
  • EV的AUC 0.840高于EI的0.731(P<.05)。
  • EI联合Marshall CT分类较单用Marshall CT分类AUC增加0.081(P=.021)。
  • EI联合A5(+1) CT分类较单用A5(+1) CT分类AUC增加0.075(P=.013)。

临床应用与阅片要点

  • 原文结论认为,整合血肿与水肿的深度学习联合模型可用于预测脑挫伤水肿的严重程度。

局限性

  • 原文未报告局限性。

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

AIM: To develop a model combing deep learning (DL) to evaluate the severity of cerebral contusion oedema.

MATERIALS AND METHODS: A DL approach along with the Tada formula and manual delineation were employed to quantify the volume of traumatic intracerebral haemorrhage (n = 141). In a separate cohort of patients with cerebral contusion oedema, participants were divided into high-risk (n = 56) and low-risk (n = 66) subgroups according to the Glasgow outcome scale at 21 days. The volume of haematoma was calculated using the aforementioned DL approach. A model integrating haematoma and oedema was developed, referred to as the oedema index (EI). Marshall computed tomography (CT) classification, A5 (+1) CT classification, oedema volume (EV), and EI were applied to evaluate the severity of oedema.

RESULTS: The volume estimation of DL approach yielded a volume percentage error of 3.59 % when compared to manual delineation. Regarding predictive efficacy for oedema severity, the area under the curve (AUC) for EV (0.840) was higher than that of the Marshall CT classification (0.664), A5 (+1) CT classification (0.692), and EI (0.731) (all P < .05). Furthermore, the combination of EI and the Marshall CT classification resulted in an AUC increase of 0.081 compared to the Marshall CT classification alone (P = .021), while the combination of EI and the A5 (+1) CT classification led to an AUC increase of 0.075 compared to the A5 (+1) CT classification alone (P = .013).

CONCLUSION: A combined model integrating haematoma and oedema using DL techniques can be used to predict the severity of cerebral contusion oedema.

来源:Clin Radiol,查看原文。版权归原作者所有。

引用

  1. Chen D, Zhan L, Shi X, et al. Severity prediction in patients with oedema in cerebral contusion using deep learning from computed tomography scans. Clin Radiol. 2025. PMID: 41240817. doi:10.1016/j.crad.2025.107113

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。