探索深度学习应用超声单视图电影在急性胆囊病变中的价值:初步结果
本研究开发基于超声单视图电影的深度学习模型,用于区分正常胆囊、非紧急胆石症和需紧急干预的急性结石性胆囊炎。纳入2017-2022年急诊右上腹痛成人患者,共186例266个电影。模型区分正常与异常影像准确率91%,区分紧急与非紧急准确率82%,识别异常特异性100%。
要点
- 深度学习模型区分正常与异常胆囊超声电影的准确率为91%。
- 模型区分需紧急干预的急性胆囊炎与非紧急情况准确率为82%。
- 模型识别异常影像的特异性为100%,无假阳性结果。
核心结论
- 仅使用易获得的单视图电影,深度学习模型在区分非紧急影像与需紧急干预的急性胆囊炎方面表现出高准确度和特异性。
研究设计与样本
- 初步研究,纳入2017-2022年急诊右上腹痛成人患者186例,共266个超声单视图电影,按最终临床诊断分为正常影像、非紧急胆石症和急性胆囊炎;电影随机分为训练集70%、验证集10%和测试集20%,并分为12帧片段进行模型训练。
关键影像表现/征象
- 正常影像组52例104个电影,非紧急胆石症组73例88个电影,急性胆囊炎组61例74个电影。
- 模型将电影分类为正常(所有片段正常)、胆石症(正常和非紧急胆石症片段)和急性胆囊炎(存在任何胆囊炎片段)。
诊断标准与关键数值
- 模型区分正常与异常影像的准确率为91%。
- 模型区分紧急(急性胆囊炎)与非紧急(胆石症或正常影像)的准确率为82%。
- 模型识别异常与正常影像的特异性为100%,无假阳性结果。
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
RATIONALE AND OBJECTIVES: In this preliminary study, we aimed to develop a deep learning model using ultrasound single view cines that distinguishes between imaging of normal gallbladder, non-urgent cholelithiasis, and acute calculous cholecystitis requiring urgent intervention.
METHODS: Adult patients presenting to the emergency department between 2017-2022 with right-upper-quadrant pain were screened, and ultrasound single view cines of normal imaging, non-urgent cholelithiasis, and acute cholecystitis were included based on final clinical diagnosis. Longitudinal-view cines were de-identified and gallbladder pathology was annotated for model training. Cines were randomly sorted into training (70%), validation (10%), and testing (20%) sets and divided into 12-frame segments. The deep learning model classified cines as normal (all segments normal), cholelithiasis (normal and non-urgent cholelithiasis segments), and acute cholecystitis (any cholecystitis segment present).
RESULTS: A total of 186 patients with 266 cines were identified: Normal imaging (52 patients; 104 cines), non-urgent cholelithiasis (73;88), and acute cholecystitis (61;74). The model achieved a 91% accuracy for Normal vs. Abnormal imaging and an 82% accuracy for Urgent (acute cholecystitis) vs. Non-urgent (cholelithiasis or normal imaging). Furthermore, the model identified abnormal from normal imaging with 100% specificity, with no false positive results.
CONCLUSION: Our deep learning model, using only readily obtained single-view cines, exhibited a high degree of accuracy and specificity in discriminating between non-urgent imaging and acute cholecystitis requiring urgent intervention.
引用
- Ge C, Jang J, Svrcek P, et al. Exploring Deep Learning Applications using Ultrasound Single View Cines in Acute Gallbladder Pathologies: Preliminary Results. Acad Radiol. 2024. PMID: 39306521. doi:10.1016/j.acra.2024.08.061
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。