文献速递胸部CT

慢性阻塞性肺疾病患者胸肌体积的CT深度学习分割及与胸肌面积的比较

Deep Learning Segmentation of Pectoralis Muscle Volume at CT and Comparison with Pectoralis Muscle Area in COPD.
PubMed 文献 · Radiol Cardiothorac Imaging · 2026-01-01

本研究为CanCOLD前瞻性队列的二次分析,开发U-Net模型在CT上分割胸肌体积(PMV),训练、内部测试和外部测试的Dice相似系数分别为0.94±0.04、0.93±0.03和0.92±0.04。与胸肌面积(PMA)相比,PMV在吸气与呼气相间的重复性更好(偏倚0.1 cm³ vs -2.7 cm²),且与FEV1、Dlco和运动峰值摄氧量的关联更强。

要点

  • U-Net模型分割胸肌体积的外部测试Dice相似系数为0.92±0.04,准确性高且可泛化。
  • PMV在吸气与呼气相间无显著偏倚(0.1 cm³,P=0.77),重复性优于PMA(-2.7 cm²,P<0.001)。
  • PMV与COPD结局指标的关联强于PMA,调整R²分别为FEV1 0.609/0.598、Dlco 0.645/0.627、VO2 0.680/0.666。

核心结论

  • 成功开发了准确且可泛化的基于CT的胸肌深度学习分割模型;与PMA相比,PMV重复性更好且与COPD结局的关联更强。

研究设计与样本

  • CanCOLD前瞻性队列研究的二次分析,数据收集于2009年11月至2015年7月;随机抽取CT扫描用于模型训练(n=96)、验证(n=16)、内部测试(n=32)及外部测试(n=32);训练U-Net模型分割PMV,以Dice相似系数评估性能,并从配对吸气与呼气扫描提取PMA和PMV评估重复性。

诊断标准与关键数值

  • 模型Dice相似系数:训练与验证集0.94±0.04,内部测试集0.93±0.03,外部测试集0.92±0.04;吸气与呼气相偏倚:PMV 0.1 cm³(P=0.77),PMA -2.7 cm²(P<0.001)。

临床应用与阅片要点

  • 原文未据此给出推荐。

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

Purpose To develop a deep learning model for segmenting pectoralis muscle volume (PMV) at CT and evaluate the reproducibility, group differences, and associations of pectoralis muscle area (PMA) and PMV with chronic obstructive pulmonary disease (COPD)-related outcomes. Materials and Methods This study was a secondary analysis of the prospective Canadian Cohort Obstructive Lung Disease study (CanCOLD, data collected from November 2009 to July 2015). Randomly sampled CT scans from CanCOLD were used for model training, validation, and internal testing (n = 96, 16, and 32, respectively) and an external dataset for external testing (n = 32). A U-Net model was trained for PMV segmentation, and performance was assessed using the Dice similarity coefficient (DSC). PMA and PMV values were extracted from paired inspiration and expiration scans to assess segmentation reproducibility. Differences between individuals with or without COPD and associations with forced expiratory volume in 1 second (FEV1), diffusing capacity of the lungs for carbon monoxide (Dlco), and peak oxygen uptake during exercise (VO2) were reported. Results Individuals included those with (n = 634; mean age, 67.3 years ± 10.1 [SD]; 394 male participants) and without (n = 601; mean age, 65.8 years ± 9.6; 327 male participants) COPD. The model yielded DSCs of 0.94 ± 0.04, 0.93 ± 0.03, and 0.92 ± 0.04 in the training and validation, internal testing, and external testing datasets, respectively. Contrary to PMV (bias, 0.1 cm3; P = .77), PMA showed bias between inspiration and expiration (bias, -2.7 cm2; P < .001). Both PMA and PMV were reduced in patients with COPD (P < .05), but PMV was more strongly associated with FEV1 (adjusted R2 [Radj2], 0.609/0.598), Dlco (Radj2, 0.645/0.627), and VO2 (Radj2, 0.680/0.666). Conclusion An accurate and generalizable CT-based deep learning model for pectoralis muscle segmentation was developed. Compared with PMA, PMV showed better reproducibility and stronger associations with COPD outcomes. Keywords: CT, Thorax, Lung, Volume Analysis, Chronic Obstructive Pulmonary Disease, Segmentation ClinicalTrials.gov identifier no. NCT00920348 © RSNA, 2026 Supplemental material is available for this article.

来源:Radiol Cardiothorac Imaging,查看原文。版权归原作者所有。

引用

  1. Genkin D, Verdawala M, Collins SÉ, et al. Deep Learning Segmentation of Pectoralis Muscle Volume at CT and Comparison with Pectoralis Muscle Area in COPD. Radiol Cardiothorac Imaging. 2026. PMID: 41784492. doi:10.1148/ryct.250060

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。