人工智能自动检测系统对不同经验医师诊断乳腺小肿块的辅助作用

目的初步探讨人工智能自动检测系统对乳腺小肿块的诊断效能以及对不同经验医师的辅助作用。 方法选取中国科学院大学宁波华美医院164个经病理证实的最大直径≤10 mm的乳腺小肿块，由4名不同经验的医师（医师A和医师B归为高年资医师组，医师C和医师D归为低年资医师组）先独立诊断，给出相应的乳腺影像报告与数据系统（BI-RADS）分类，结果设为A1组、B1组、C1组和D1组。4周后，4名医师再次联合应用人工智能自动检测系统诊断，结果设为A2组、B2组、C2组和D2组。人工智能自动检测系统诊断结果设为M组。以病理结果为金标准，计算不同组医师诊断的敏感度、特异度、准确性、阴性预测值和阳性预测值，绘制受试者操作特征曲线。采用Kappa检验比较不同组观察者间的一致性。 结果病理结果显示，164个乳腺肿块中良性117个（71.34%），恶性47个（28.66%）。人工智能自动检测系统对乳腺小肿块有良好的诊断效能，敏感度、特异度、准确性分别为91.49%、90.6%、90.85%。联合应用人工智能自动检测系统后高年资医师的诊断效能有上升趋势，敏感度、特异度分别为A1组82.98%、82.05%；A2组87.23%、89.74%；B1组80.85%、84.62%；B2组85.11%、89.74%。低年资医师在人工智能自动检测系统的辅助下诊断效能明显提高，敏感度、特异度分别为C1组76.60%、74.36%；C2组82.98%、82.05%；D1组68.09%、73.50%；D2组80.85%、80.34%。此外低年资医师借助人工智能自动检测系统对乳腺小肿块BI-RADS分类的观察者间一致性明显提高，低年资医师间的Kappa值由0.236提高到0.549，低年资医师与高年资医师的Kappa值由0.268~0.284提高到0.432~0.540。 结论人工智能自动检测系统对乳腺小肿块良恶性的判断具有较高的诊断效能，其对不同经验医师的辅助作用不同，对低年资医师的影响大于高年资医师。人工智能自动检测系统有助于提高低年资医师BI-RADS分类观察者间的一致性。

Role of an automated artificial intelligence detection system in diagnosis of small breast masses by physicians with varying levels of experience

ObjectiveTo preliminarily investigate the diagnostic efficacy of an automated artificial intelligence (AI) detection system for small breast masses, as well as its complementary role for physicians with varying levels of experience. MethodsA total of 164 pathologically confirmed small breast masses with a diameter of up to 10 mm were chosen at Hwa Mei Hospital, University of Chinese Academy of Sciences, and four physicians (doctors A and B were senior physicians, and doctors C and D were junior ones) with varying levels of experience independently diagnosed the masses according to the Breast Imaging Reporting and Data System (BI-RADS) classification, and the results were reported as groups A1, B1, C1, and D1. Four weeks later, the four physicians again jointly applied the AI automatic detection system to diagnose the masses, and the results were reported as groups A2, B2, C2, and D2. The results of the AI automatic detection system were reported as group M. Taking pathological results as the gold standard, the sensitivity, specificity, accuracy, negative predictive value, and positive predictive value of diagnosis by physicians in different groups were calculated, and the receiver operating characteristic curve was drawn. The Kappa test was used to calculate the interobserver agreement between groups. ResultsPathological results showed that among the 164 breast masses, 117 (71.34%) were benign and 47 (28.66%) were malignant. The AI automatic detection system had good diagnostic efficacy for small breast lumps, with a sensitivity, specificity, and accuracy of 91.49%, 90.6%, and 90.85%. The diagnostic efficacy of senior physicians after the joint application of the AI automatic detection system tended to increase, with a sensitivity and specificity of 82.98% and 82.05% in group A1; 87.23% and 89.74% in group A2; 80.85% and 84.62% in group B1; and 85.11% and 89.74% in group B2. The diagnostic efficacy of senior physicians with the aid of the AI automatic detection system was significantly improved, with a sensitivity and specificity of 76.60% and 74.36 in group C1; 82.98% and 82.05% in group C2; 68.09% and 73.50% in group D1; and 80.85% and 80.34% in group D2. The interobserver agreement of BI-RADS classification of small breast masses by low senior physicians with the help of the AI automatic detection system was significantly higher. The Kappa value among junior physicians increased from 0.236 to 0.549. The Kappa value among junior and senior physicians increased from 0.268-0.284 to 0.432-0.540. ConclusionThe AI automated detection system has high diagnostic efficacy in determining the benignity and malignancy of small breast masses, and its adjunctive effect varies among experienced physicians, with a greater impact on junior physicians than on senior physicians. The AI automated detection system helps to improve the interobserver agreement of BI-RADS classification among junior physicians.