2022 年第 3 期 第 17 卷

人工智能在乳腺钼靶X线摄影诊断乳腺疾病中的应用研究

Application of artificial intelligence in mammography diagnosis of breast diseases

作者：刘东婷1黄秀娟2温兆赢1石春彦1王文川1徐磊1刘家祎1

英文作者：Liu Dongting1 Huang Xiujuan2 Wen Zhaoying1 Shi Chunyan1 Wang Wenchuan1 Xu Lei1 Liu Jiayi1

单位：1首都医科大学附属北京安贞医院医学影像科，北京100029；2河北省大厂回族自治县人民医院放射科，大厂065300

英文单位：1Department of Medical Imaging Beijing Anzhen Hospital Capital Medical University Beijing 100029 China； 2Department of Radiology Dachang Hui Autonomous County People′s Hospital Hebei Province Dachang 065300 China

关键词：乳腺癌；乳腺钼靶X线；人工智能

英文关键词：Breastcancer;Mammography;Artificialintelligence

• 摘要：

• 目的 探讨人工智能在乳腺钼靶X 线摄影诊断乳腺疾病中的应用价值。方法  选取2018年2月至2021年5月于首都医科大学附属北京安贞医院行乳腺钼靶X线摄影检查并进行病理诊断的73例女性患者的临床资料进行回顾性分析。使用人工智能对乳腺钼靶X线摄影检查病灶进行诊断，并与放射科医师诊断结果进行比较，分析二者在乳腺钼靶X线摄影诊断中的一致性。结果 73例患者共采集146例次乳腺钼靶X线摄影图像，其中127例次有病理检测结果。Kappa一致性检验结果显示，医师、人工智能及医师+人工智能3种乳腺钼靶X线摄影诊断方式均与病理检测结果呈现一致性（均P＜0.001）。医师+人工智能诊断的敏感度（94.55%）、特异度（97.22%）、总符合率（96.06%）、约登指数（0.92）及阴性预测值（95.89%）最高。146例次乳腺钼靶X线摄影图像中，医师与人工智能在诊断良性钙化、可疑病变钙化、不对称结构扭曲、增大淋巴结及肿块或结节数目方面呈一致性（均P<0.001）。此外，医师观察图像发现10例患者乳头凹陷和乳腺皮肤增厚，人工智能未体现这2项内容。人工智能明确显示乳腺影像报告和数据系统（BI-RADS）分级的图像共76例次，将这些图像交于医师进行BI-RADS分级,结果显示医师与人工智能对BI-RADS分级呈一致性(P＜0.001)。结论  人工智能在乳腺钼靶X线摄影诊断乳腺病变方面的表现与放射科医师相似，可提高影像科医师对乳腺癌筛查和诊断的准确性。

• Objective   To investigate the application value of artificial intelligence (AI) in mammography diagnosis of breast diseases. Methods   From February 2018 to May 2021, clinical data of 72 female patients undergoing mammography and pathological diagnosis in Beijing Anzhen Hosptial, Capital Medical University were retrospectively analyzed. AI was used to diagnose the lesions in mammography, the results were compared with those of radiologists, and the concordance between them in mammography was analyzed. Results  Totally 146 times of mammographic images were collected from 73 patients, and 127 times of which had pathologic results. Kappa consistency test showed that the mammography diagnosis of radiologists, AI and radiologists+AI was concordance with pathologic results (all P<0.001). Radiologists+AI diagnosis had the best sensitivity (94.55%), specificity (97.22%), overall concordance rate (96.06%), Youden index (0.92) and negative predictive value (95.89%). Among 146 times of mammographic images, radiologists was concordance with AI in diagnosis of benign calcifications, calcifications in suspicious lesions, asymmetric architectural distortion, enlarged lymph nodes and the number of masses or nodules (all P<0.001). In addition, radiologists found nipple dimpling and breast skin thickening in 10 patients in mammographic images, while those were not found by AI. AI clearly showed 76 times of mammographic images of breast imaging reporting and data system (BI-RADS) grading, and those were referred to the radiologists for BI-RADS grading. It was showed that radiologists was concordance with AI in BI-RADS grading (P<0.001). Conclusions The performance of AI in mammography diagnosis of breast lesions is similar to that of radiologists, which is helpful for radiologists to improve the accuracy of screening and diagnosis of breast cancer.