基于深度学习实现上腹部超声标准切面识别与多器官分割
本研究基于单中心回顾性数据,利用465例上腹部超声视频中提取的5535帧标准切面图像,开发并评估两阶段深度学习框架。先以卷积神经网络完成12类上腹部标准切面识别,再对已识别切面进行肝脏、胆囊、胰腺、脾脏等多器官分割。结果显示切面识别模型测试集准确率达97.00%,器官分割中胰腺表现最弱。
要点
- 标准切面识别模型测试集精确率97.08%、准确率97.00%,可作为超声质控与智能导航的前置步骤。
- 肝脏、胆囊、右肾分割表现较强,胰腺分割测试IoU仅53.76%,阅片与报告时需警惕胰腺区域自动分割误差。
- 该模型依赖单一中心真实世界数据,外推至其他设备或人群前需行外部验证。
核心结论
- 在上腹部超声单中心真实世界数据中,标准切面识别与多器官分割的两阶段深度学习方法可行,两项任务互补。
- 胰腺分割性能最低,提示超声图像中胰腺边界模糊及个体差异对分割模型构成挑战。
研究设计与样本
- 回顾性研究,使用单中心真实世界临床超声数据,共采集521例上腹部超声检查视频,465例符合纳入标准。
- 由超声医师提取并审核标准切面帧,共纳入5535幅图像,覆盖肝脏、胆囊、胰腺和脾脏的12类临床定义上腹部标准切面。
- 采用两阶段序贯框架:先以卷积神经网络进行标准切面识别作为前置筛选步骤,再由深度学习分割网络仅对已识别标准切面行器官分割。
关键影像表现/征象
- 切面识别模型在测试集上精确率97.08%、召回率97.00%、特异度99.70%、准确率97.00%,各数据子集表现稳定。
- 标准切面约束下的器官分割对肝脏、胆囊和右肾结果较强,对脾脏表现较低,对胰腺表现最弱。
- 分割性能模式在验证集与测试集间保持一致,未见明显过拟合驱动的结果分化。
诊断标准与关键数值
- 标准切面识别测试集性能:Precision 97.08%,Recall 97.00%,Specificity 99.70%,Accuracy 97.00%。
- 胰腺分割测试集IoU为53.76%,为多器官中最低值。
临床应用与阅片要点
- 可将标准切面识别模型作为超声图像质控、扫描导航和报告自动化的上游步骤,优先识别肝、胆、脾、胰相关12类切面。
- 自动分割对肝、胆囊和右肾较可靠,可用于预勾画或测量辅助;胰腺分割结果需人工复核后使用。
局限性
- 病例来自单一中心,样本量465例,未行外部验证,模型普适性尚不能确定。
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 develop and evaluate deep learning models for upper abdominal ultrasound standard section recognition and downstream multi-organ segmentation.
METHODS: This retrospective study used real-world clinical ultrasound data from a single center. A total of 521 upper abdominal ultrasound examination videos were collected, of which 465 met the inclusion criteria. Standard section frames were extracted and reviewed by sonographers, yielding 5535 images representing 12 clinically defined upper abdominal standard sections across the liver, gallbladder, pancreas, and spleen. A two-stage sequential framework was implemented, in which standard section recognition using a convolutional neural network served as a prerequisite screening step, followed by organ segmentation using a deep learning-based segmentation network applied only to recognized standard sections. Data were divided into training, validation, and test sets with strict case-level separation.
RESULTS: The standard section recognition model demonstrated stable performance across datasets, achieving Precision 97.08%, Recall 97.00%, Specificity 99.70%, and Accuracy 97.00% on the test set. Organ segmentation performed under standard section constraints yielded strong results for the liver, gallbladder, and right kidney, whereas performance was lower for the spleen and particularly the pancreas (test IoU 53.76%). Performance patterns were consistent across the validation and test subsets.
CONCLUSION: Deep learning models for standard section recognition and multi-organ segmentation are feasible within this single-center real-world dataset and address two complementary tasks in upper abdominal ultrasound; external validation is still needed to determine broader generalizability.
引用
- Wang, Zhang, Xie, et al. Deep Learning-Based Standard Section Recognition and Multi-Organ Segmentation in Upper Abdominal Ultrasound. Ultrasound Med Biol. 2026. PMID: 42399171. doi:10.1016/j.ultrasmedbio.2026.06.003
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。