研究目的

以北京市顺义区乳腺癌筛查项目进行典型案例分析，评估我国基于超声的乳腺癌筛查项目的实施效果以及时间变化趋势，并分析现存瓶颈问题；同时基于全国多中心随机对照研究，初步评估人工智能辅助超声技术应用于我国乳腺癌筛查的效果。

研究方法

1. 北京顺义区乳腺癌筛查项目实施现状评估：对2016—2023年北京顺义区35～64岁筛查人群的个案数据进行回顾性分析，计算顺义区近五轮筛查覆盖率、乳腺癌检出率及变化趋势，同时分别评估超声、乳腺X线环节的漏诊情况以及乳腺癌检出与检出轮次的关系。

2. 人工智能辅助筛查技术在中国应用效果初步探索：2024年4月—12月在中国14个现场招募35～64岁女性为研究对象，采用随机对照研究方法，评估将基于手持超声实时动态图像的人工智能技术应用于我国一般人群筛查中的乳腺癌检出效果及筛查效果，并通过医生接触AI技术时间、筛查医生年资与专业、超声分辨率、妇女年龄、BMI等因素分层进行敏感性分析。此外，基于AI辅助超声探索超声4A类妇女管理路径。

研究结果

1. 北京顺义区乳腺癌筛查项目实施现状评估：2016—2023年顺义区乳腺癌累计筛查覆盖率由42.44%上升至69.88%，差异有统计学意义（Z=158.39，P<0.001），2016—2021年周期内覆盖率由42.44%下降至39.97%，差异有统计学意义（Z=-14.74，P<0.001）。乳腺X线遵从率为45.72%～58.72%，呈现上升趋势（Z=24.85，P<0.001）。总活检遵从率为35.77%。乳腺癌检出率由0.83‰上升至1.61‰，差异有统计学意义（Z=4.03，P<0.001），早诊率均高于75%。超声阴性（BI-RADS 1类、2类）检出乳腺癌占比降低，差异有统计学意义（Z=-3.20，P<0.001），但超声BI-RADS 0类、3类、4类、5类阳性预测值保持在较低水平。乳腺X线阴性检出乳腺癌占比提高，差异有统计学意义（Z=2.78，P=0.006）。

2. 人工智能辅助筛查技术在中国应用效果初步探索：2024年4月至2024年12月全国14个现场共有58 795名35～64岁（平均年龄：49.35±7.90）女性被纳入分析集。AI辅助组超声0类、3类、4类、5类比例显著高于常规组（8.02% vs 6.94%，P<0.001）。以超声4类、5类为阳性时，AI辅助组乳腺癌检出率、灵敏度、阳性预测值及AUC值均略低于常规组，但差异无统计学意义（P均>0.05）。以超声BI-RADS 4B及以上为阳性时，AI辅助组灵敏度、AUC值均显著高于常规组（50.00% vs 20.51%，P=0.006；0.750 vs 0.602，P=0.004）。在医生接触AI技术时间分层下，接触时间的第1四分位数时AI辅助组AUC值高于常规组（0.812 vs 0.562，P=0.024），差异有统计学意义；到接触时间的第4四分位数时，AI辅助组AUC值与常规组相近（0.750 vs 0.667，P=0.530）。以超声0类、3类、4类、5类为阳性时，AI辅助组乳腺癌检出率略高于常规组，差异无统计学意义（1.36‰ vs 1.30‰，P=0.836），但阳性预测值低于常规组，差异无统计学意义（1.69% vs 1.87%，P=0.662）。超声4A类妇女阳性预测值较低（9.17%），且妇女活检失访率较高（62.78%）。在超声4A类并接受乳腺X线检查者的子集中模拟超声4A补充MAM新策略（超声4A类转乳腺X线，乳腺X线3类及以上者转乳腺病理检查），新策略相较于现行策略（超声4A类均转乳腺病理检查）可在不漏诊的条件下减少36.21%的不必要活检，其中AI辅助组减少41.67%不必要活检，常规组减少27.27%的不必要活检。

研究结论

1. 北京顺义区乳腺癌筛查项目实施现状评估：顺义区周期内筛查覆盖率呈现下降趋势。乳腺癌检出率逐年提高，早诊率处于较高水平。超声假阳性比例及乳腺X线漏诊比例呈现上升趋势，且进一步检查依从率有待提高，尤其是乳腺病理活检，提示当地应加强初筛机构及转诊机构超声及乳腺X线检查准确性，优化异常人群管理机制。未来应重点关注多年未参加筛查的妇女，探索应用新兴筛查技术如人工智能筛查技术、自动容积超声以提升筛查的精准度，同时尝试打破信息壁垒，建立高效、及时的筛查随访体系。

2. 人工智能辅助筛查技术在中国应用效果初步探索：实时动态超声AI技术可提高医生对超声BI-RADS 4B类及以上病例的识别能力，并且这种能力提升可随着AI使用时间而强化。实时动态超声AI技术显著增加了异常人群转诊比例，且当前基线数据尚未观察到转诊人群中乳腺癌检出比例的显著增加，未来还需要通过随访数据进一步评估其应用于我国乳腺癌筛查的价值。聚焦超声4A类妇女患癌风险较低的现状，超声4A类补充MAM的模拟策略可在不影响乳腺癌检出率的情况下降低此类妇女的不必要活检率，其中AI辅助技术相对减少的不必要活检比例更大。未来可基于真实世界研究，进一步探索该策略的应用效果及可行性，逐步优化我国筛查策略，提高乳腺癌筛查服务质量。

Objectives

A typical case study of the Shunyi Breast Cancer Screening Project was conducted to assess the implementation effect of the ultrasound-based breast cancer screening programme in China, as well as the trend of time change, and to analyse the existing obstacles; at the same time, based on the national multi-centre randomized controlled study, the effect of the application of the AI-assisted ultrasound technology in breast cancer screening was preliminarily assessed in China.

Materials and methods

1.Evaluation of the implementation status of the breast cancer screening program in Shunyi: Case data from the screened population aged 35-64 years in Shunyi, from 2016-2023, were retrospectively analyzed to calculate the coverage rate of the last five screening rounds, the detection rate of breast cancer, and the trend of change, as well as to assess the leakage of ultrasound and mammography sessions and the association of breast cancer detection and detection rounds, respectively.

2.Preliminary exploration of the application effect of artificial intelligence-assisted screening technology in China: Women aged 35-64 years were recruited as study subjects at 14 sites in China from April to December 2024. Through establishing a randomized controlled study, the screening performance of applying AI technology based on real-time dynamic images of handheld ultrasound to general population screening in China was assessed. And the sensitivity analysis was conducted according to the factors such as the time of physician exposure to AI technology, screening doctors' professional level, ultrasound resolution, age and BMI. In addition, the management pathways for women in ultrasound category 4A were explored based on AI-assisted ultrasound.

Results

1.Evaluation of the implementation status of the breast cancer screening program in Shunyi: The cumulative breast cancer screening coverage rate in Shunyi increased from 42.44% to 69. 88% in 2016-2023, with a statistically significant difference in comparison (Z=158.39, P<0. 001), and the coverage rate in each round decreased from 42. 44% to 39. 97% in the 2016-2021, with a difference of statistically significant (Z=-14.74, P<0. 001). Mammography compliance rates increased from 45.72% to 58.72%, with a statistically significant difference (Z=24.85, P<0. 001). Total biopsy compliance rates were 35.77%. Breast cancer detection rate increased from 0.83‰ to 1.61‰, with a statistically significant difference (Z=4.03, P<0. 001), and both early diagnosis rates were >75%. The proportion of breast cancers detected negatively by ultrasound decreased and the difference was statistically significant (Z=-3.20, P<0.001), but the positive predictive value of ultrasound categories 0, 3, 4 and 5 remained low. The percentage of negatively detected breast cancers on mammograms increased and the difference was statistically significant (Z=2.78, P=0. 006).

2.Preliminary exploration of the application effect of artificial intelligence-assisted screening technology in China: A total of 58,795 women aged 35-64 years (49.35±7.90) from 14 sites were included in the analysis set from April 2024 to December 2024.The proportion of BI-RADS categories 0, 3, 4, and 5 was significantly higher in the AI-assisted group than in the conventional group (8.02% vs 6.94%, P<0.001). Considering BI-RADS 4 and 5 categories of ultrasound as positive, the detection rate, sensitivity, positive predictive value and AUC value in the AI-assisted group were slightly lower than those in the conventional group, but the differences were not statistically significant (all P>0.05). Regarding ultrasound BI-RADS 4B and above as positive, the sensitivity and AUC values were significantly higher in the AI-assisted group than in the conventional group (50.00% vs 20.51%, P=0.006; 0.750 vs 0.602, P=0.004). Under the time stratification of physician exposure to AI technology, the initial AUC value of 0.812 in AI-assisted group was higher than that of the conventional group at 0.562 (P=0.024), and by the 4th quartile of exposure time, the AI-assisted group AUC value was similar to that of the conventional group (0.750 vs 0.667, P=0.530). When using ultrasound categories 0, 3, 4, and 5 as positive, the breast cancer detection rate was slightly higher in the AI-assisted group than in the conventional group, with a non-statistically significant difference (1.36‰ vs. 1.30‰, P=0.836), but the positive predictive value was lower than in the conventional group, with a statistically non-significant difference (1.69% vs. 1.87%, P=0.662). Women with ultrasound category 4A had a lower positive predictive value (9.17%) and women had a highly biopsy dropout rate (62.78%). Among women with ultrasound results of BI-RADS 4A who also underwent mammography, a new strategy of combining ultrasound 4A with MAM was simulated. Compared with the current strategy (all BI-RADS 4A cases are referred to pathological biopsy), the new strategy could reduce unnecessary biopsies by 36.21% without missed diagnoses. And the reduction was 41.67% in the AI-assisted group and 27.27% in the conventional group.

Conclusions

1.Evaluation of the implementation status of the breast cancer screening program in Shunyi: The cycle screening coverage rate in Shunyi shows a decreasing trend. The detection rate of breast cancer increased year by year, and the early diagnosis rate was at a high level. The proportion of ultrasound false positives and mammogram missed diagnoses showed an increasing trend. The compliance rate of further examination needs to be improved, especially breast pathology biopsy, which suggests that the local government should strengthen the accuracy of ultrasound and mammogram examination in the primary screening and referral institutions, and optimize the management mechanism of abnormal populations. In the future, we should focus on women who have not participated in screening for many years, explore the application of new screening technologies such as artificial intelligence screening technology and automatic volumetric ultrasound to improve the accuracy of screening, and try to break down the information barriers to establish an efficient and timely screening follow-up system.

2.Preliminary exploration of the application effect of artificial intelligence-assisted screening technology in China: Real-time dynamic ultrasound AI technology improves physicians' ability to recognize ultrasound BI-RADS 4B and higher cases, and this improvement intensifies over time with AI use. However, real-time dynamic ultrasound AI technology significantly increased the proportion of referrals for further examinations, and the baseline data have not yet observed a significant increase in the proportion of breast cancer detection in the referral population. In the future, it is necessary to further evaluate the effect of AI-assisted technology used in breast cancer screening through follow-up data. Focusing on the status of ultrasound category 4A women with lower cancer risk, the simulation strategy of ultrasound category 4A supplemented with MAM can reduce the unnecessary biopsy rate, without affecting the breast cancer detection rate. And the relative reduction of unnecessary biopsy by AI-assisted technology is greater. In the future, the effect and feasibility of this management strategy should be further explored based on real-world studies to gradually optimize our screening strategy and improve the quality of breast cancer screening.