基于深度学习的慢性胰腺炎患者CT胰腺自动分割
本研究开发了基于人工智能的胰腺分割模型,用于在常规CT中精确勾画慢性胰腺炎患者的复杂胰腺,并用内部和外部测试集验证。模型在奥尔堡测试集Dice为0.85±0.08,在卑尔根数据集为0.79±0.19,在美国国立卫生研究院数据集为0.79±0.18。内脏脂肪和胰腺体积与Dice评分正相关(r=0.45和r=0.53,均p<0.0001),CT参数无显著影响。
要点
- 模型在慢性胰腺炎和健康受试者中均获得较高分割精度,跨站点和扫描仪表现稳健。
- 胰腺体积和内脏脂肪面积与分割Dice评分正相关(r分别为0.53和0.45)。
- CT参数对模型性能无显著影响(所有p值>0.07)。
核心结论
- 人工智能模型在慢性胰腺炎患者和健康受试者的胰腺分割中表现出高精度和稳健性,跨不同站点和扫描仪,提示其具有临床应用的潜力。
研究设计与样本
- 多中心研究,纳入550例回顾性收集的CT扫描,来自奥尔堡(n=373;224例慢性胰腺炎,150例健康受试者)和卑尔根医院(n=97例慢性胰腺炎),以及美国国立卫生研究院在线数据集(n=80,健康受试者);奥尔堡数据集分为训练队列(n=326)和内部测试集(n=47),卑尔根和NIH数据集作为外部测试集;AI模型采用nnU-Net架构,以Sørensen-Dice指数评估性能。
诊断标准与关键数值
- 胰腺分割AI模型在奥尔堡测试集Dice评分为0.85±0.08。
- 模型在卑尔根数据集Dice评分为0.79±0.19。
- 模型在NIH数据集Dice评分为0.79±0.18。
- 内脏脂肪和胰腺体积与Dice评分正相关(r=0.45和r=0.53,均p<0.0001)。
- CT参数对模型性能无显著影响(所有p值>0.07)。
临床应用与阅片要点
- 模型可用于慢性胰腺炎患者和健康受试者常规CT扫描中的胰腺自动分割。
- 胰腺体积和内脏脂肪面积与分割精度正相关,阅片时需注意这些因素可能影响模型表现。
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
OBJECTIVES: This study aimed to develop an artificial intelligence (AI)-based segmentation model for accurate delineation of the complex pancreas in patients with chronic pancreatitis (CP) using computer tomography (CT) scans obtained during routine clinical work-up. Validation was performed with internal and external test datasets. A secondary objective was to evaluate the impact of visceral fat area (at the third lumbar level), pancreas volume, and CT parameters on model performance.
METHODS: This multicenter study included 550 retrospectively collected CT scans from Aalborg (n = 373; 224 CP, 150 healthy subjects) and Bergen Hospitals (n = 97 CP), and an online dataset from the National Institutes of Health (NIH) (n = 80, healthy subjects). The Aalborg dataset was divided into a training cohort (n = 326) and an internal test set (n = 47), while the Bergen and NIH datasets served as external test sets. The AI model employed the nnU-Net architecture, with performance evaluated using the Sørensen-Dice index. Correlations with visceral fat, pancreas volume, and CT parameters were assessed.
RESULTS: The pancreas segmentation AI model achieved a Dice score of 0.85 ± 0.08 on the Aalborg test set, 0.79 ± 0.19 on the Bergen dataset, and 0.79 ± 0.18 on the NIH dataset. Visceral fat and pancreas volume positively correlated with Dice scores (r = 0.45 and r = 0.53, both p < 0.0001), while CT parameters had no significant impact (all p-values > 0.07).
CONCLUSION: The AI model demonstrated high accuracy and robustness in pancreas segmentation of both CP patients and healthy subjects, and across diverse sites and scanners, suggesting its potential for clinical application.
引用
- Nalliah S, Mark EB, Liedenbaum MH, et al. Deep-learning based automated pancreas segmentation on CT scans of chronic pancreatitis patients. Eur J Radiol. 2025. PMID: 40408911. doi:10.1016/j.ejrad.2025.112175
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。