探讨基于深度学习技术的MammoWorks?乳腺健康智能检测系统在乳腺肿瘤检测中的应用价值。

本回顾性研究收集2015年1月至2017年4月期间就诊于陕西省肿瘤医院、乳腺X线BI-RADS 5～6级的患者448例，均手术治疗且临床病理资料齐全。另外用随机数字表法收集同期参加健康体检乳腺X线检查提示BI-RADS 1级的215例正常人群作为对照。以上全部研究对象乳腺X线影像学资料经MammoWorks?乳腺检测系统分析，以患者的病理结果及正常人群的2年随访结果为金标准，分析MammoWorks?乳腺检测系统检测乳腺肿瘤的敏感度、特异度、准确率、阳性预测值、阴性预测值、阳性似然比、阴性似然比及每幅图的假阳性标记数等。率的比较使用χ²检验及Fisher确切概率法，每幅图假阳性标记数比较使用非参数检验(Kruskal-Wallis检验)，并使用Kappa检验评价组间变量结果的一致性。

总计纳入663例女性，X线摄片2 652张。总计识别病灶2 284个，真阳性标记929个，假阳性病灶1 355个。真阳性病例333例，真阴性病例126例，假阳性病例89例，假阴性病例115例。MammoWorks?分析敏感度为74.3%(333/448)，特异度为58.6%(126/215)，阳性预测值为78.9%(333/422)，阴性预测值为52.3%(126/241)，准确率为69.2%(459/663)，阳性似然比为1.80，阴性似然比为0.44。每幅图假阳性标记数为0.50(0.00～0.75)。2种拍摄体位(头尾位和内外侧斜位)下，MammoWorks?系统检测效能差异有统计学意义(Kappa＝0.278，P<0.001)。在不同年龄、BI-RADS分级、肿瘤部位、病理分期、病理类型、分子分型的患者中，MammoWorks?检测效能差异无统计学意义(χ²＝3.341、1.056、7.103、8.911、5.170、7.803，P均>0.050)；在不同乳腺密度、病灶类型、肿瘤直径的患者中，MammoWorks?检测效能差异有统计学意义(χ²＝7.985、15.543、18.652，P均<0.050)。每幅图片假阳性标记数仅在不同乳腺密度分组中差异有统计学意义(χ²＝15.024，P<0.050)。

基于深度学习技术的MammoWorks?乳腺健康智能检测系统在乳腺肿瘤的辅助诊断中具有一定的应用价值，但其检测效能仍需进一步提高。

To investigate the value of MammoWorks? system based on deep learning technology in breast tumor detection.

We enrolled 448 patients with breast lesions at X-ray BI-RADS grade 5-6 in Shaanxi Provincial Tumor Hospital from January 2015 to April 2017. All patients underwent operation, with complete clinical and pathological data. Additionally, using a random number table, 215 healthy people who had physical examination in our hospital at the same period (X-ray BI-RADS grade 1) were randomly enrolled as control and all of them had no breast diseases in the two-year follow-up. The mammographic data of all subjects were analyzed by MammoWorks? system. With the pathological results of patients and the 2-year follow-up results of healthy people as the gold standard, the sensitivity, specificity, accuracy, positive predictive value, negative predictive value, positive likelihood ratio and negative likelihood ratio of MammoWorks? system in the detection of breast tumors were analyzed, as well as the number of false-positive marks in one mammograph. The rate comparison was performed using χ² test and Fisher exact probability method. The number of false-positive marks in one mammograph was compared using nonparametric test (Kruskal-Wallis test) and Kappa test was used to evaluate the consistency of the results between different subgroups.

Totally 2 652 X-ray photographs from 663 females were analyzed. A total of 2 284 lesions were marked, including 929 true-positive and 1 355 false-positive. There were 333 cases of true-positive, 126 true-negative, 89 false-positive and 115 false-negative. The sensitivity of MammoWorks? system was 74.3%(333/448), specificity 58.6%(126/215), positive predictive value 78.9%(333/422), negative predictive value 52.3%(126/241), accuracy 69.2%(459/663), positive likelihood ratio 1.80 and negative likelihood ratio 0.44. The number of false-positive marks in one mammograph was 0.50(0.00~0.75). The sensitivity of MammoWorks? system showed a significant difference between craniocaudal (CC) view and mediolateral oblique (MLO) view of X-ray (Kappa=0.278, P<0.001). The detection efficiency of MammoWorks? system presented no significant difference in patients with different age, BI-RADS grade, tumor location, pathological stage, pathological type and molecular type (χ²=3.341, 1.056, 7.103, 8.911, 5.170, 7.803, P>0.050), while the detection efficiency of MammoWorks? system was significantly different in patients with different breast density, lesion type, and tumor diameter (χ²=7.985, 15.543, 18.652, P<0.050). The number of false-positive marks in one mammograph presented a significant difference in patients with different breast density (χ²=15.024, P<0.050).

Based on deep learning technology, the MammoWorks? system is helpful in the auxiliary diagnosis of breast tumors, but its detection efficiency still needs to be improved.