| 摘要: |
| 目的 探讨基于钙化特征参数的深度学习模型对乳腺导管原位癌（DCIS）的诊断价值。方法 回顾性分析2016年1月至2022年12月经南方医科大学深圳医院或中山大学肿瘤防治中心乳腺钼靶X线检查发现钙化的患者资料，根据病理诊断结果将患者分为良性病变组569例、DCIS组263例。采用深度学习建立钙化特征检测和分类模型，筛选差异有统计学意义的特征参数，并采用ROC曲线分析各特征参数及深度学习模型、传统机器学习模型对DCIS的诊断效能。结果 20个特征参数中线样分支状钙化数、细颗粒状钙化数、段样分布率、簇状分布率、种群密度等5个参数在良性病变组与DCIS组之间差异有统计学意义，以DCIS组数值较高（P均＜0.05）。线样分支状钙化数、细颗粒状钙化数、段样分布率、簇状分布率、种群密度识别DCIS的ROC曲线下面积（AUC）分别是0.762、0.732、0.725、0.757、0.810，灵敏度分别为81.2%、85.9%、80.1%、87.8%、86.4%，特异度分别是76.0%、71.6%、70.3%、73.4%、63.7%。通过这5个特征参数组合建立的深度学习模型识别DCIS的AUC值、灵敏度、特异度分别为0.823、89.3%、77.3%，高于独立特征参数和传统机器学习模型支持向量机、K-邻近算法、线性判别分析、logistic回归模型（AUC分别为0.771、0.801、0.765、0.734，灵敏度分别为79.7%、80.9%、77.8%、74.4%，特异度分别为57.8%、62.9%、56.6%、47.9%）。结论 筛选诊断价值较高的钙化特征并建立深度学习模型有助于提高乳腺X线片上DCIS的诊断识别水平。 |
| 关键词: 乳腺肿瘤 深度学习 乳房X线摄影术 钙化 导管原位癌 |
| DOI：10.16781/j.CN31-2187/R.20220231 |
| 投稿时间：2022-03-18修订日期：2022-07-08 |
| 基金项目:深圳市宝安区科技计划基础研究项目（2021JD041），广东省医学科学技术研究基金项目（A2017222），广东省大学生创新训练计划项目（S202112121097）. |
|
| Diagnostic value of deep learning model based on calcification recognition for breast ductal carcinoma in situ |
| WANG Jinhua1,LIU Lizhi2,XU Cui1,LUO Min1,LIU Yubao1* |
(1. Medical Image Center, Shenzhen Hospital of Southern Medical University, Shenzhen 518000, Guangdong, China;
2. Diagnostic Imaging and Intervening Center, Cancer Center, Sun Yat-sen University, Guangzhou 510060, Guangdong, China
*Corresponding author) |
| Abstract: |
| Objective To explore the diagnostic value of deep learning model based on calcification characteristic parameters for breast ductal carcinoma in situ (DCIS). Methods From Jan. 2016 to Dec. 2022, data of patients with calcification found by mammography in Shenzhen Hospital of Southern Medical University or Cancer Center of Sun Yat-sen University were retrospectively analyzed. According to the pathological diagnosis results, the patients were divided into benign group (569 cases) and DCIS group (263 cases). The calcification feature detection and classification model were established by deep learning, and the characteristic parameters with statistically significant differences were screened. The receiver operating characteristic curve (ROC) was used to analyze the diagnostic effectiveness of each characteristic parameter, the deep learning model and traditional machine learning models on DCIS. Results Among the 20 characteristic parameters, 5 (the number of linear branching-like calcification, the number of granular calcification, rate of segment distribution, rate of clustered distribution, and population density) had significant differences between the 2 groups, and the DCIS group had higher values (all P＜0.05). The area under curve (AUC) values of ROC in the number of linear branching-like calcification, the number of granular calcification, rate of segment distribution, rate of clustered distribution and population density for identifying DCIS were 0.762, 0.732, 0.725, 0.757 and 0.810, respectively, with sensitivities of 81.2%, 85.9%, 80.1%, 87.8% and 86.4%, and specificities of 76.0%, 71.6%, 70.3%, 73.4% and 63.7%, respectively. The AUC value, sensitivity and specificity of the deep learning model established by the combination of the 5 feature parameters for recognition of DCIS were 0.823, 89.3% and 77.3%, respectively, which were higher than those of independent feature parameters and traditional machine learning models, including support vector machine, K-nearest neighbor, linear discriminant analysis, and logistic regression (AUC values were 0.771, 0.801, 0.765, and 0.734, sensitivities were 79.7%, 80.9%, 77.8%, and 74.4%, and specificities were 57.8%, 62.9%, 56.6%, and 47.9%, respectively). Conclusion Screening the calcification features with high diagnostic value and establishing a deep learning model are helpful to the diagnosis of DCIS with mammograms. |
| Key words: breast neoplasms deep learning mammography calcification ducal carcinoma in situ |
