摘要
目的 探讨基于深度学习的乳腺X线辅助诊断(DL)系统对乳腺钙化检出和良恶性分类的临床价值。方法 回顾性分析在2020年1月~2022年12月在徐州市中心医院接受双侧乳腺X线检查的400例患者的头尾位和内外斜位影像资料。以2位具有15年以上乳腺X线诊断经验的副主任医师对乳腺钙化的一致判断作为标准组,由1位低年资住院医师、1位高年资主治医师和DL系统分别盲法独立阅片,经过4周洗脱期后,由联合模型(低年资医师+DL系统)再次盲法独立阅片。结合双向表χ2检验,评价不同乳腺ACR类型、钙化形态和分布、BI-RADS分类对钙化检出的影响,并采用ROC曲线下面积(AUC)评价低年资住院医师、高年资主治医师、DL系统和联合模型(低年资住院医师+DL系统)对可疑钙化检出的性能差异。结果 1600幅图像(400例患者)共检出BI-RADS 3级及以上可疑钙化975处。低年资住院医师A,高年资主治医师B、DL系统和联合模型对钙化检出的敏感度分别为81.95%、96.62%、93.03%、96.41%。高年资主治医师B、DL系统和联合模型对钙化检出的敏感度不受乳腺ACR类型、钙化形态和分布、BI-RADS分类影响,而低年资住院医师A对钙化检出的敏感度受其影响。联合模型(低年资住院医师+DL系统)在预测钙化良恶性方面具有良好的AUC值、敏感度和特异性,分别为0.891、90.0%和88.2%,和低年资住院医师之间存在差异(P<0.01)。在DL系统帮助下,低年资住院医师的诊断性能得到明显改善,AUC值由0.740提升到0.891。结论 DL系统对BI-RADS 3级及以上可疑钙化检出敏感度高且具有较高的良恶性钙化分类性能,与高年资主治医师相当。在DL系统的帮助下,低年资医师可以减少钙化漏诊、误诊,提高乳腺癌筛查和诊断的准确性。
Objective To investigate the clinical value of the deep learning-based mammography-assisted diagnosis(DL)system for breast calcification detection and benign and malignant classification.Methods A retrospective analysis was performed on the craniocaudal and internal and external oblique imaging data of 400 patients who underwent bilateral mammography in Xuzhou Central Hospital from January 2020 to December 2022.The unanimous judgment of two associate chief physicians with more than 15 years of experience in mammography diagnosis was used as the standard group,the images were blinded and independently reviewed by 1 junior resident,1 senior attending physician,and the DL system,respectively.After a 4-week washout period,the images were blinded and independently reviewed by the combined model(junior resident+DL system)again.Combined with two-way table chi-square test,the effects of different ACR types,morphology and distribution of calcification,and BI-RADS classification on the detection of calcification were evaluated.The area under the curve(AUC)was used to evaluate the difference in the detection of suspicious calcification among junior residents,senior attending physician,DL system and combined model(junior resident+DL system).Results A total of 975 suspicious calcifications of BI-RADS3 grade and above were detected in 1600 images(400 patients).The sensitivities of junior resident A,senior attending physician B,DL system and combined model were 81.95%,96.62%,93.03%and 96.41%,respectively.The sensitivity of senior attending physician B,DL system and combined model to calcification detection was not affected by breast ACR type,morphology and distribution of calcification,and BI-RADS classification,while the sensitivity of junior resident A was affected by it.The combined model(junior resident+DL system)had high AUC value,sensitivity and specificity in predicting the benign and malignant nature of calcifications,with 0.891,90.0%and 88.2%,respectively,which differed from that of the junior resident(P<0.01).With the
作者
翟天旭
张敏伟
张子秋
孔德懿
李德春
ZHAI Tianxu;ZHANG Mingwei;ZHANG Ziqiu;KONG Deyi;LI Dechun(Affiliated Xuzhou Clinical College of Xuzhou Medical University,Xuzhou 221009,China;Department of Radiology,Xuzhou Central Hospital,Xuzhou 221009,China)
出处
《分子影像学杂志》
2024年第1期25-30,共6页
Journal of Molecular Imaging
基金
江苏省十四五医学重点学科项目(ZDXK202237)
徐州市科学技术局社会发展项目(KC15SH061)。
关键词
乳腺X线摄影
可疑钙化
深度学习
乳腺癌
人工智能
mammography
suspicious calcification
deep learning
breast cancer
artificial intelligence
