基于乳腺超声动态连续影像的深度学习模型建立与初步应用

目的 探讨基于乳腺超声动态连续影像的深度学习模型建立方法并对其效能进行初步验证。方法 对506例女性进行乳腺超声扫查，存储实时动态图像，导入深睿影像智能分析平台，采用基于深度学习的端到端的肿块检出网络对原始动态序列图像进行分析提取，训练建立最优化深度学习模型，并对模型的效能进行测试验证，数据采用Python3.6软件进行统计分析。结果 单帧乳腺超声影像的肿块检出敏感率（0.1、0.2、0.5/scan）为76.6%、84.2%、86.0%，序列乳腺超声影像的肿块检出敏感率（0.1、0.2、0.5/scan）为77.3%、91.8%、95.3%；0.1/scan，单帧乳腺超声影像的肿块检出与序列乳腺超声影像的肿块检出无统计学意义（P >0.05），0.2/scan，单帧乳腺超声影像的肿块检出与序列乳腺超声影像的肿块检出有统计学意义（P <0.05），0.5/scan, 单帧乳腺超声影像的肿块检出与序列乳腺超声影像的肿块检出有统计学意义（P <0.05）。结论 基于乳腺超声动态连续影像的深度学习模型能提高乳腺超声影像的肿块检出率。

Establishment and preliminary application of deep learning model based on dynamic continuous breast ultrasound images of breast ultrasounde

objectives: Aimed to establish and validate deep learning model based on dynamic continuous image of mammary gland. Methods: 506 cases of female breast ultrasound were performed, real-time dynamic images were stored, and the images were imported into the intelligent analysis platform of deep learning. The end-to-end tumor detection network based on deep learning was used to analyze and extract the original dynamic sequence images, and the optimized deep learning model was trained and established, and the validity of the model was tested and verified. Data were analyzed using Python 3.6 software. Results: The sensitivity of single frame breast ultrasound images (0.1, 0.2, 0.5/scan) were 76.6%, 84.2%, 86.0%, and the sensitivity of sequential breast ultrasound images (0.1, 0.2, 0.5/scan) were 77.3%, 91.8%, 95.3%. At 0.1/scan, there was no statistical significance in lump detection between single frame breast ultrasound image and serial breast ultrasound image(P>0.05), At 0.2/scan, there was statistical significance in lump detection between single frame breast ultrasound image and serial breast ultrasound image (P<0.05), At 0.5/scan, there was statistical significance in lump detection between single frame breast ultrasound image and serial breast ultrasound image (P<0.05). Conclusions: Deep learning model based on dynamic continuous breast ultrasound image can improve the breast tumor detection rate.