查看更多>>摘要：目的 探讨基于超声造影视频的深度学习模型鉴别肝肿瘤良恶性的临床价值。 方法 回顾性收集2010年5月至2022年6月西南医科大学附属医院因肝脏肿物接受超声造影检查的1 213例患者，以2021年12月31日为时间截点，将入组患者分为训练集和独立测试集。训练集采用TimeSformer算法为基础架构，利用视频时间维度的滑动窗口从超声造影动脉期视频中获取多个定长时间片段，将多段视频的特征融合后得出整个视频的分类结果，从而构建基于超声造影视频的深度学习模型。独立测试集采用ROC曲线验证模型的有效性，并将模型与三名不同肝脏超声造影年资超声医师(R1、R2和R3，分别有3、6和10年肝脏超声造影经验)进行比较。 结果 研究共纳入1 213例患者的1 213个肝脏肿物，其中训练集1 066例(恶性426例)，独立测试集147例(恶性50例)。基于超声造影的深度学习模型在训练集的曲线下面积(AUC)为0.93±0.01，在独立测试集的AUC为0.89±0.01，准确性、敏感性、特异性、阳性预测值(PPV)和阴性预测值(NPV)分别为80.42%、74.19%、92.00%、94.52%和65.71%。三名医师中，R1的诊断性能最差，其准确性、敏感性、特异性、PPV和NPV分别为67.83%、51.61%、98.00%、97.96%和52.13%，而R3的上述指标分别为82.52%、76.36%、94.00%、95.95%和68.12%。McNemar′s检验显示，R1与深度学习模型的差异有统计学意义(P<0.001)，而R2和R3与深度学习模型的差异无统计学意义(P＝0.720、0.868)。此外，模型对单个病例的分析时间为(340.24±16.32)ms，而医师对单个病例分析的平均时间为62.9 s。 结论 基于超声造影的深度学习模型可较好地区分肝脏肿物的良恶性，有望达到与经验较丰富超声医师相当的诊断水平。 Objective To investigate the clinical value of deep learning model based on contrast enhanced ultrasound (CEUS) video in the differential diagnosis of benign and malignant liver tumors. Methods Between May 2010 and June 2022, 1 213 patients who underwent CEUS examination for liver masses in the Affiliated Hospital of Southwest Medical University were retrospectively collected, and the enrolled patients were divided into training and independent test cohorts with December 31, 2021 as the time cut-off. In the training cohort, the TimeSformer algorithm was used as the infrastructure, and multiple fixed-time segments were obtained from CEUS arterial videos by using the sliding window of the video, and the classification results of the entire video were obtained after fusing the features of multiple segments, so as to build a deep learning model based on CEUS videos. In the independent test cohort, ROC curves were used to verify the validity of the model and compared with three radiologists with different CEUS experience (R1, R2, and R3, with 3, 6, and 10 years of CEUS experience, respectively). Results A total of 1 213 patients with liver masses were included in the study, including 1 066 patients in the training cohort (426 cases of malignancy) and 147 patients in the independent test cohort (50 cases of malignancy). The area under curve (AUC)value of deep learning model was 0.93±0.01 in the training cohort and 0.89±0.01 in the independent test cohort, and the accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were 80.42%, 74.19%, 92.00%, 94.52% and 65.71%, respectively. Among the three radiologists, R1 had the lowest diagnostic performance, with accuracy, sensitivity, specificity, PPV and NPV of 67.83%, 51.61%, 98.00%, 97.96% and 52.13%, respectively, while the above indicators of R3 were 82.52%, 76.36%, 94.00%, 95.95% and 68.12%, respectively. McNemar′s test showed that the difference between R1 and the deep learning model was statistically significant (P<0.001), while the differences between R2 and R3 and the deep learning model were not statistically significant (P=0.720, 0.868). In addition, the analysis time of the model for a single case was (340.24±16.32)ms, while the average analysis time of radiologists was 62.9 s. Conclusions The deep learning model based on CEUS can better identify benign and malignant liver masses, and may reach the diagnostic level of experienced radiologists.
万方数据