人工智能筛查乳腺癌的效率有多高？比普通医师快30倍！

　　乳腺癌是发生在乳腺腺上皮组织的恶性肿瘤，一直是威胁女性健康的恐怖杀手之一。此前在该疾病的筛查上，医师们多数采用乳腺摄片的检验方法，但这种方式通常需要乳腺活组织检查来辅助，这就为女性带来了不必要的痛苦。不过，未来人工智能（AI）将大幅降低乳腺疾病检测为女性带来的不适。

　　2016年8月29日，美国癌症学会官方期刊《癌症》在线发表美国休斯敦卫理公会医院研究所癌症中心、纽约康奈尔大学医学院、德克萨斯大学公共卫生学院、圣安东尼奥健康科学中心、德国慕尼黑技术大学的研究报告，开发出一款AI软件，该软件在解析乳腺摄片时比普通医师快30倍，其准确率更高达99%，可以直观的将X光图转译成诊断信息，方便医师快速对患者病情作出判断，以免耽误病情。

　　为了检验这款软件的实力，该研究收集了543位乳腺癌患者的乳腺摄片和病理组织切片报告，还安排了各种相关医学发现试图迷惑AI。几个小时之内，AI就成功完成了任务。而另一边，两位乳腺癌病理医师却花了50～70小时才完成50位患者的诊断。

　　来自美国疾病预防控制中心（CDC）和美国癌症学会（ACS）的数据显示，每年美国大约有1210万人接受乳腺摄片监测，但是其中差不多有一半可能为“假阳性”，这就造成了每年160万人左右为了求安心接受乳腺活检，而其中20%的女性根本就没病。

　　除了减少患者痛苦和节省医师时间，该软件还能制住不断攀升的抗癌成本。美国国家癌症研究所（NCI）预计，到2020年，国家花在癌症上的钱将达到史无前例的1580亿美元。

　　根据调查机构GrandView的数据显示，虽然活检结果正确率只有60%～80%，但是乳腺活检设备的市场规模到2024年将达9.11亿美元。约翰霍普金斯大学研究人员也表示，每年浪费在最简单的非浸润性乳腺癌上的资金就高达3500万美元。

　　研究人员对AI软件充满信心，希望该软件能更好地帮助医师评估患者的病情，并减少患者活检的痛苦。

Cancer. 2016 Aug 29. [Epub ahead of print]

Correlating mammographic and pathologic findings in clinical decision support using natural language processing and data mining methods.

Patel TA, Puppala M, Ogunti RO, Ensor JE, He T, Shewale JB, Ankerst DP, Kaklamani VG, Rodriguez AA, Wong ST, Chang JC.

Houston Methodist Cancer Center, Houston, Texas; Houston Methodist Research Institute, Houston, Texas; Weill Cornell Medicine, New York, New York; Houston Methodist Hospital, Houston, Texas; University of Texas School of Public Health, Houston, Texas; University of Texas Health Science Center at San Antonio, San Antonio, Texas; Technical University of Munich, Garching, Germany.

BACKGROUND: A key challenge to mining electronic health records for mammography research is the preponderance of unstructured narrative text, which strikingly limits usable output. The imaging characteristics of breast cancer subtypes have been described previously, but without standardization of parameters for data mining.

METHODS: The authors searched the enterprise-wide data warehouse at the Houston Methodist Hospital, the Methodist Environment for Translational Enhancement and Outcomes Research (METEOR), for patients with Breast Imaging Reporting and Data System (BI-RADS) category 5 mammogram readings performed between January 2006 and May 2015 and an available pathology report. The authors developed natural language processing (NLP) software algorithms to automatically extract mammographic and pathologic findings from free text mammogram and pathology reports. The correlation between mammographic imaging features and breast cancer subtype was analyzed using one-way analysis of variance and the Fisher exact test.

RESULTS: The NLP algorithm was able to obtain key characteristics for 543 patients who met the inclusion criteria. Patients with estrogen receptor-positive tumors were more likely to have spiculated margins (P = .0008), and those with tumors that overexpressed human epidermal growth factor receptor 2 (HER2) were more likely to have heterogeneous and pleomorphic calcifications (P = .0078 and P = .0002, respectively).

CONCLUSIONS: Mammographic imaging characteristics, obtained from an automated text search and the extraction of mammogram reports using NLP techniques, correlated with pathologic breast cancer subtype. The results of the current study validate previously reported trends assessed by manual data collection. Furthermore, NLP provides an automated means with which to scale up data extraction and analysis for clinical decision support.

KEYWORDS: data mining; imaging characteristics; mammographic to pathologic correlation; natural language processing; subtypes of breast cancer

PMID: 27571243

DOI: 10.1002/cncr.30245