超声影像组学与深度学习预测膀胱癌HER2状态
回顾性纳入190例膀胱尿路上皮癌患者,比较MedSAM半自动与手动分割超声图像后构建的影像组学、深度迁移学习及联合模型对HER2状态的预测效能。结果显示半自动分割Dice为0.824±0.109,其联合模型测试集AUC最高为0.842。
要点
- 半自动分割联合模型测试集AUC达0.842,优于手动分割的0.805。
- MedSAM半自动分割Dice为0.824±0.109,接近人工勾画水平。
- 随机森林算法被选用于构建影像组学与深度学习影像组学标签。
核心结论
- 基于MedSAM的半自动分割性能接近人工专家,结合临床、影像组学与深度学习特征的联合模型在预测膀胱尿路上皮癌HER2状态方面表现良好。
研究设计与样本
- 回顾性研究,纳入2022年1月至2025年5月共190例膀胱尿路上皮癌患者,训练集133例、测试集57例。
- 对膀胱超声图像分别采用MedSAM半自动分割与手动分割勾画感兴趣区,提取影像组学特征并训练ResNet18提取深度迁移学习特征。
- 影像组学特征与深度学习特征融合构建深度学习影像组学模型,通过单因素与多因素分析筛选临床、病理及超声特征并联合建模。
诊断标准与关键数值
- MedSAM半自动分割的Dice相似系数为0.824±0.109。
- 半自动分割组联合模型测试集AUC为0.842(0.734-0.951)。
- 半自动分割组深度学习影像组学模型测试集AUC为0.835(0.725-0.945)。
- 半自动分割组影像组学模型测试集AUC为0.814(0.700-0.930),深度迁移学习模型为0.794(0.673-0.915)。
- 手动分割组联合模型测试集AUC为0.805(0.693-0.917),影像组学模型为0.766(0.638-0.894),深度迁移学习模型为0.800(0.681-0.920),深度学习影像组学模型为0.792(0.676-0.908)。
临床应用与阅片要点
- 在日常阅片中可借助MedSAM半自动分割膀胱超声病灶,提高勾画一致性与效率。
- 半自动分割的联合模型对HER2状态预测效能略优,可作为膀胱尿路上皮癌HER2无创评估的参考工具。
局限性
- 本研究为回顾性单中心设计,样本量有限,结果需进一步验证。
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: To evaluate semi-automatic segmentation of bladder ultrasound images and develop human epidermal growth factor receptor 2 (HER2) prediction models using radiomics and deep learning.
METHODS: We retrospectively analyzed patients with urothelial carcinoma of the bladder from January 2022 to May 2025. The regions of interest (ROIs) were delineated using both semi-automatic segmentation based on segment anything in medical images (MedSAM) and manual segmentation. Radiomic features were extracted and screened to construct radiomics models. ResNet18 was trained to extract deep transfer learning (DTL) features, and the reduced-dimensional features were employed to build DTL models. The radiomic features and DL features were combined to develop deep learning radiomics (DLR) models. Univariate analysis and multivariate analysis were conducted to identify clinical, pathological, and ultrasonic features, and the selected feature was integrated with the most suitable DLR model to develop a combined model.
RESULTS: A total of 190 patients were enrolled (133 in the train set and 57 in the test set). The MedSAM-based semi-automatic segmentation achieved a dice similarity coefficient score of 0.824 ± 0.109. The random forest algorithm was selected for constructing both the radiomics signature and the DLR signature. In the manual segmentation group, the Rad, DTL, DLR, and combined signature achieved an area under curve (AUC) of 0.766 (0.638 - 0.894), 0.800 (0.681 - 0.920), 0.792 (0.676 - 0.908), and 0.805 (0.693 - 0.917) in the test set, respectively. In the semi-automatic segmentation group, the Rad, DTL, DLR, and combined signature achieved an AUC of 0.814 (0.700 - 0.930), 0.794 (0.673 - 0.915), 0.835 (0.725 - 0.945), and 0.842 (0.734 - 0.951) in the test set, respectively.
CONCLUSION: MedSAM-based semi-automatic segmentation achieves segmentation performance approaching human experts. The model combining clinical, radiomics, and deep learning features showed promising performance in predicting HER2 status in urothelial carcinoma of the bladder.
引用
- Zhuang, Xie, Chen, et al. Radiomics and Deep Learning Models Based on Manual and Semi-automatic Segmentation of Ultrasound Images for HER2 Prediction in Bladder Cancer. Acad Radiol. 2026. PMID: 42760168. doi:10.1016/j.acra.2026.08.101
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。