多模态卷积神经网络评估胎兔肺发育:B型超声与剪切波弹性成像
本研究基于167只胎兔共1670张B型超声与剪切波弹性成像(SWE)图像,构建单模态及多模态卷积神经网络模型,以肺组织学分期为参考标准预测胎肺发育。结果显示SWE单模态模型总体准确率(95.9%)优于B型(86.5%)及多模态模型(87.9%),提示SWE有望成为临床评估胎儿肺成熟度的无创工具。
要点
- SWE单模态模型总体准确率最高,提示弹性信息对胎肺分期价值更大。
- 肺泡期多模态模型特异度(98.6%)与阳性预测值(97.3%)略优于SWE单模态。
- B型模型对肺泡期AUC仅0.670,单独用于胎肺成熟评估可靠性有限。
核心结论
- 基于SWE的卷积神经网络模型可预测胎兔肺发育阶段,可能成为临床评估胎儿肺成熟度的非侵入性工具。
研究设计与样本
- 动物实验研究,共167只胎兔(孕23至30天),获取1670张超声胎肺图像(含B型与SWE)。
- 剖宫产后测量体重、Apgar评分、间质肺比值及肺泡灌洗液二棕榈酰磷脂酰胆碱水平。
- 以胎肺组织学分类(小管期、囊状期、肺泡期)为参考标准,构建基于B型、SWE及两者联合的卷积神经网络模型。
关键影像表现/征象
- SWE单模态模型总体准确率为95.9%,高于B型模型86.5%及多模态模型87.9%。
- SWE单模态模型在预测小管期和囊状期的大多数指标上优于B型及多模态模型(p < 0.05)。
- 肺泡期中SWE与多模态模型的AUC均为0.890,明显高于B型模型的0.670(p < 0.05)。
诊断标准与关键数值
- SWE单模态模型:总体准确率95.9%;肺泡期准确率96.2%、特异度96.4%、阳性预测值93.5%、AUC 0.890。
- B型单模态模型:总体准确率86.5%;肺泡期准确率86.8%、特异度79.6%、阳性预测值72.7%、AUC 0.670。
- 多模态模型:总体准确率87.9%;肺泡期准确率95.9%、特异度98.6%、阳性预测值97.3%、AUC 0.890。
临床应用与阅片要点
- SWE提供的组织弹性信息比单纯B型灰度信息更能反映胎肺成熟过程中组织微结构变化。
- 肺泡期评估中多模态模型在排除假阳性方面略有优势,可用于提高阳性报告的信心。
局限性
- 本研究为胎兔动物实验,周期与影像特征与人类胎儿存在差异,尚不能直接外推至临床应用。
- 参考标准依赖剖宫产后组织学分类及生化指标,与孕期实时预测之间存在时间差。
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: To evaluate the performance of convolutional neural network (CNN)-based models for predicting fetal rabbit lung development: unimodal models using B-mode and shear wave elastography (SWE) and a multimodal model combining B-mode and SWE.
METHODS: A total of 1670 ultrasound fetal lung images (B-mode and SWE) were acquired from 167 fetal rabbits (23-30 d gestation). Post-cesarean, body weight, Apgar scores, interstitial lung ratios and alveolar lavage dipalmitoylphosphatidylcholine levels were measured. The developed CNN models, based on fetal lung histological classification (canalicular, saccular and alveolar stages), extracted features from B-mode or SWE images to predict fetal lung development.
RESULTS: The SWE unimodal model (95.9%) showed superior total accuracy over the B-mode (86.5%) and multimodal models (87.9%) and outperformed them on most metrics for predicting the canalicular and saccular stages (p < 0.05). For the alveolar stage, SWE (96.2%, 96.4%, 93.5%, 0.890) and multimodal models (95.9%, 98.6%, 97.3%, 0.890) outperformed B-mode (86.8%, 79.6%, 72.7%, 0.670) in accuracy, specificity, positive predictive value (PPV) and area under the curve (AUC) (p < 0.05), with the multimodal model showing a slight advantage in specificity (98.6%) and PPV (97.3%).
CONCLUSION: The SWE-based models can predict fetal rabbit lung development and may serve as a promising non-invasive tool for clinical assessment of fetal lung maturity.
引用
- Huang, Zhang, Wang, et al. Multimodal Convolutional Neural Network Model for Evaluating Development of Fetal Rabbit Lung Using B-mode and Shear Wave Elastography Ultrasound Images. Ultrasound Med Biol. 2026. PMID: 41486052. doi:10.1016/j.ultrasmedbio.2025.12.003
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。