【摘要】 目的：评估基于全卷积单阶段神经网络的人工智能（AI）诊断系统在乳腺X线良恶性病变鉴别诊断中的应用价值。方法：回顾性收集288例患者的乳腺X线图像及病理学资料，由AI系统及高、中、低年资医师分别对乳腺良恶性病变进行乳腺影像报告与数据系统（BI-RADS）分类。以病理结果为金标准，计算诊断准确率、阳性预测值、阴性预测值，使用受试者工作特征（ROC）曲线分析其诊断效能。比较AI与不同年资医师诊断准确性的差异，并比较不同乳腺密度、病灶类型及病灶大小对诊断的影响。结果：针对288个病灶（良性病灶100个、恶性病灶188个），AI及高、中、低年资医师诊断乳腺良性病变的准确率分别为89.00%、94.00%、86.00%、73.00%，诊断恶性病变的准确率分别为84.04%、90.96%、86.70%、71.81%；阳性预测值分别为93.49%、96.61%、92.09%、83.33%；阴性预测值分别为74.79%、84.68%、77.48%、57.94%;ROC曲线下面积（AUC）值分别为0.87、0.92、0.86、0.72。良性病变中，AI与低年资医师诊断准确度差异有统计学意义（P<0.05）；恶性病变中，AI与高、低年资医师诊断准确度差异有统计学意义（P<0.05）。对于不同乳腺密度、病灶类型及病灶大小，AI及高、中、低年资医师诊断准确度差异均无统计学意义（均P>0.05）。结论：AI诊断系统在乳腺X线良恶性病变鉴别诊断中准确率较高，明显高于低年资医师，与中年资医师相当，低于高年资医师，且不受乳腺密度、病灶类型及病灶大小的影响，具有较好的临床应用价值。更多还原

【Abstract】 Purpose: To evaluate the application value of artificial intelligence(AI) diagnostic system based on fully convolutional one-stage neural network in the differential diagnosis of benign and malignant lesions on mammography. Methods: The mammographic images and pathological data of 288 patients were retrospectively collected. The breast lesions were classified using Breast Imaging Reporting and Data System(BI-RADS) by AI system and radiologists(senior, middle and junior seniority, respectively).The diagnostic accuracy, positive predictive value and negative predictive value were calculated using pathological results as the gold standard, and the diagnostic efficacy was analyzed using the receiver operating characteristics(ROC) curve. The differences in diagnostic accuracy between AI and radiologists with different seniority were compared. The impact on diagnosis by different breast densities, lesional types and lesional sizes were also analyzed. Results: Of all the 288 lesions(100 benign lesions and 188 malignant lesions), the diagnostic accuracy of AI and senior, middle, and junior radiologists in distinguishing benign breast lesions were 89.00%, 94.00%, 86.00% and 73.00%, respectively, while the diagnostic accuracy for malignant lesions were 84.04%, 90.96%, 86.70% and 71.81%, respectively. The positive predictive values were 93.49%, 96.61%, 92.09% and 83.33%, respectively, while the negative predictive values were 74.79%, 84.68%, 77.48% and 57.94%, respectively. The area under ROC curve(AUC) values were 0.87, 0.92, 0.86 and 0.72, respectively. There was statistically significant difference in the diagnostic accuracy on benign lesions between AI and junior radiologists(P<0.05). In malignant lesions, the differences of diagnostic accuracy between AI and radiologists with senior and junior seniority were statistically significant(P<0.05). There was no statistically significant difference in diagnostic accuracy between AI and radiologists with different seniority between different breast densities, lesion types and lesion sizes(all P>0.05). Conclusion: AI system has high accuracy in the differential diagnosis of benign and malignant breast lesions on mammography, which is significantly higher than that of junior radiologists but lower than that of senior radiologists. Moreover, the diagnostic accuracy of AI is not affected by breast density, lesional type and lesional size.更多还原