文献速递腹部技术与安全USMR

基于超声造影的深度学习模型提升肝内胆管癌的检测

Improving Detection of Intrahepatic Cholangiocarcinoma with a Contrast-enhanced US-based Deep Learning Model.
PubMed 文献 · Radiol Imaging Cancer · 2025-01-01

本回顾性研究利用49个中心的超声造影检查数据训练并验证了基于四种算法的深度学习模型,用于诊断肝内胆管癌。外部测试集B中,模型AUC达0.91,与高年资超声造影医师和MRI医师表现相似,并显著提升中低年资医师的诊断性能。

要点

  • 超声造影深度学习模型诊断肝内胆管癌的AUC为0.91,与高年资医师相当。
  • 模型辅助可显著提升中低年资超声造影医师的诊断性能。
  • ResNet-50算法在外部测试集A中表现最佳(AUC 0.92)。

核心结论

  • 基于超声造影的深度学习模型诊断性能与高年资超声造影医师相似,并能提升中低年资超声造影医师的诊断性能。

研究设计与样本

  • 回顾性研究,训练集804例、验证集344例超声造影检查来自49个中心,外部测试集A来自2个独立中心共153例,外部测试集B来自51个中心共240例,以AUC评估诊断性能。

诊断标准与关键数值

  • ResNet-50在外部测试集A中AUC为0.92,变异系数5.1;外部测试集B中模型AUC 0.91,低年资超声造影医师0.72,中年资医师0.78,高年资超声造影医师0.87,高年资MRI医师0.89。

临床应用与阅片要点

  • 模型辅助后,低年资超声造影医师AUC从0.72升至0.89,中年资医师从0.78升至0.90,均与高年资超声造影医师表现相似。

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 (DL) model based on contrast-enhanced US (CEUS) to help radiologists diagnose intrahepatic cholangiocarcinoma (iCCA). Materials and Methods In this retrospective study (July 2017-December 2023), CEUS examinations from 49 centers were used to train and validate a DL model using four algorithms (BNInception, MobileNet-v2, ResNet-50, and VGG-19). External test set A, collected from two independent centers, was used to evaluate model performance. External test set B, collected from 51 centers, was used to compare the DL model's performance on iCCA diagnosis with that of three CEUS radiologists and one MRI radiologist and to assess the effect of DL-assisted interpretation on radiologist performance. Diagnostic performance was assessed using the area under the receiver operating characteristic curve (AUC). Results A total of 1148 CEUS examinations were divided into training (n = 804) and validation (n = 344) sets. External test sets A and B included 153 (mean age, 55.52 years ± 11.12 [SD]; 120 male patients) and 240 (mean age, 55.03 years ± 11.25; 184 male patients) CEUS examinations, respectively. Among the four evaluated algorithms, ResNet-50 achieved the best performance (AUC, 0.92) and robustness (coefficient of variation, 5.1) in external test set A. In external test set B, the DL model achieved a higher AUC than did the junior (0.91 vs 0.72, P < .01) and midlevel (0.91 vs 0.78, P < .01) CEUS radiologists and performance similar to that of the senior CEUS radiologist (0.91 vs 0.87, P = .32) and senior MRI radiologist (0.91 vs 0.89, P = .56). With DL assistance, diagnostic performance of the junior and midlevel CEUS radiologists improved significantly (from 0.72 to 0.89 [P < .01] and from 0.78 to 0.90 [P < .01], respectively), reaching performance similar to that of the senior CEUS radiologist (P = .50 for junior radiologist and P = .94 for midlevel radiologist). Conclusion A CEUS-based DL model demonstrated diagnostic performance similar to that of a senior CEUS radiologist and improved the performance of junior and midlevel CEUS radiologists. Keywords: Applications-Ultrasound, Deep Learning, Ultrasound-Contrast, Abdomen/GI, Liver, Oncology ClinicalTrials.gov identifier no. NCT04682886 Supplemental material is available for this article. © RSNA, 2025.

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

引用

  1. Ding W, Li B, Zhao L, et al. Improving Detection of Intrahepatic Cholangiocarcinoma with a Contrast-enhanced US-based Deep Learning Model. Radiol Imaging Cancer. 2025. PMID: 41236388. doi:10.1148/rycan.250078

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。