Chen Xiao, Zhang Yang, Zhou Jiejie, Pan Yong, Xu Hanghui, Shen Ying, Cao Guoquan, Su Min-Ying, Wang Meihao
Department of Radiology, First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China (X.C., J.Z., Y.P., Y.S., G.C., M.W.).
Department of Radiological Sciences, University of California, Irvine, CA (Y.Z., J.Z., M-Y.S.); Department of Radiation Oncology, University of California, Irvine, CA (Y.Z.).
Acad Radiol. 2025 Mar;32(3):1287-1296. doi: 10.1016/j.acra.2024.10.031. Epub 2024 Nov 3.
Detection and diagnosis of architectural distortion (AD) on digital breast tomosynthesis (DBT) is challenging. This study applied artificial intelligence (AI) using deep learning (DL) algorithms to detect AD, followed by radiomics for classification.
500 cases with AD on DBT reports were identified; the earlier 292 cases for training, and the later 208 cases for testing. The DL Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to automatically localize abnormalities and generate a region of interest (ROI), which was put into the radiomics model to estimate the malignancy probability for constructing ROC curves. Radiologists delineated ROI manually for comparison. Cases were categorized into pure AD and AD associated with other features, including mass, regional high-density, and calcifications. The ROC curves were compared using the DeLong test.
The overall malignancy rate was 57% (285/500). Of them, 267 cases were classified as pure AD, and the malignancy rate (106/267 = 39.7%) was significantly lower compared to AD cases associated with other features (179/233 = 76.8%, p < 0.01). In the testing set, the diagnostic AUC was 0.82 when using the manual ROI and 0.84 when using the DL-generated ROI. In the more challenging pure AD cases, DL-generated ROI yielded an AUC of 0.77, significantly lower than 0.86 for AD associated with other features.
DL could detect AD on DBT, and the diagnostic performance was comparable to manual ROI. The strategy worked for pure AD, but the performance was worse than that for AD with other features.
在数字乳腺断层合成（DBT）上检测和诊断结构扭曲（AD）具有挑战性。本研究应用基于深度学习（DL）算法的人工智能（AI）来检测AD，随后进行放射组学分类。
确定了500例DBT报告中有AD的病例；较早的292例用于训练，较晚的208例用于测试。应用DL梯度加权类激活映射（Grad-CAM）自动定位异常并生成感兴趣区域（ROI），将其放入放射组学模型中以估计恶性概率，从而构建ROC曲线。放射科医生手动勾勒ROI进行比较。病例分为单纯AD和与其他特征相关的AD，其他特征包括肿块、区域高密度和钙化。使用DeLong检验比较ROC曲线。
总体恶性率为57%（285/500）。其中，267例被分类为单纯AD，其恶性率（106/267 = 39.7%）显著低于与其他特征相关的AD病例（179/233 = 76.8%，p < 0.01）。在测试集中，使用手动ROI时诊断AUC为0.82，使用DL生成的ROI时为0.84。在更具挑战性的单纯AD病例中，DL生成的ROI的AUC为0.77，显著低于与其他特征相关的AD的0.86。
DL可在DBT上检测AD，其诊断性能与手动ROI相当。该策略对单纯AD有效，但性能比与其他特征相关的AD更差。
