基于18F-FDG PET/CT的深度学习算法鉴别肺结节良恶性的研究

18F-FDG PET/CT-based deep learning methods for the differentiation of malignant and benign pulmonary nodules

    摘要
  • 摘要:
    目的  探讨基于18F-氟脱氧葡萄糖（FDG） PET/CT的深度学习算法鉴别肺结节良恶性的效能。
    方法  回顾性分析2017年1月至2021年12月于华中科技大学同济医学院附属同济医院行18F-FDG PET/CT的223例肺结节患者的临床资料，其中男性137例、女性86例，中位年龄56岁，范围26~82岁。按8∶2的比例将患者采用sklearn.model-selection.train_test_split函数随机分为训练集和验证集。分别以1.0、1.5、2.0作为PET摄取增高掩模的标准化摄取值（SUV）临界值，基于PET/CT影像构建用于鉴别肺结节良恶性的多模态融合注意力网络。结合深度学习特征和临床特征构建综合模型。不符合正态分布的计量资料采用Mann-Whitney U检验进行比较；计数资料采用卡方检验进行比较；采用多因素Logistic回归分析分析变量与组织病理学类型的关系；采用受试者工作特征（ROC）曲线计算模型的诊断效能。
    结果 最大标准化摄取值（SUVmax）（P<0.001）、癌胚抗原（CEA）水平（P=0.026）、性别（P<0.001）是肺结节良恶性的独立影响因素。基于多模态融合注意力网络的诊断效能优于单模态注意力网络曲线下面积（AUC）：0.780 对 0.763，准确率：0.800 对0.739，灵敏度：0.833 对0.767，特异度：0.733 对 0.688，F1-分数：0.847 对 0.793。SUV=1.5在PET摄取增高掩模与肺结节CT掩模共同引导下的多模态融合注意力网络的鉴别诊断效能最好（AUC：0.831，准确率：0.844）。综合影像与临床特征结合可以进一步提高肺结节良恶性的诊断效能（AUC：0.864，准确率：0.822）。
    结论 基于18F-FDG PET/CT的深度学习模型能够有效鉴别肺结节的良恶性，综合模型有助于进一步提高诊断效能。

     

    Abstract:
    Objective  To explore the efficacy of deep learning methods based on 18F-fluorodeoxyglucose (FDG) PET/CT in distinguishing benign and malignant pulmonary nodules.
    Methods  Retrospective analysis of clinical data of 223 patients with pulmonary nodules who underwent 18F-FDG PET/CT at Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology from January 2017 to December 2021, including 137 males and 86 females, with a median age of 56 years and an age range from 26 to 82 years. Patients were randomly divided into the training set and test set at a ratio of 8∶2 using sklearn.model_selection.train_test_split function. Using 1.0, 1.5, and 2.0 as the standardized uptake value (SUV) cut-off values for PET hypermetabolic masks, a multi-modal fusion attention network was constructed based on PET/CT images for distinguishing between benign and malignant pulmonary nodules. A comprehensive model was also constructed by combining deep learning features and clinical features. The Mann-Whitney U test was used to compare the differences between measurement data that do not conform to normal distribution. Count data was compared using Chi square test. Multivariate Logistic regression analysis was conducted to investigate the relationship between variables and histopathological types. Diagnostic efficiency was evaluated using receiver operating characteristic curves.
    Results  Maximum standardized untake value (SUVmax) (P<0.001), carcinoma embryonic antigen (CEA) levels (P=0.026), and gender (P<0.001) were identified as independent risk factors for the classification of lung nodules. The diagnostic efficacy of the multi-modal fusion attention network was superior to that of the single-modal attention network (area under curve (AUC): 0.780 vs. 0.763, accuracy: 0.800 vs. 0.739, sensitivity: 0.833 vs. 0.767, specificity: 0.733 vs. 0.688, and F1-score: 0.847 vs. 0.793). The differential diagnostic performance of the multi-modal fusion attention network guided by both PET uptake enhancement mask and pulmonary nodule CT mask is best when SUV=1.5, achieving an AUC of 0.831 and an accuracy of 0.844. The combination of imaging and clinical features further improved the diagnostic efficacy (AUC of 0.864 and accuracy of 0.822).
    Conclusions  The deep learning model based on 18F-FDG PET/CT effectively distinguished benign and malignant pulmonary nodules. Furthermore, the comprehensive model further optimized the diagnostic efficacy.

     

    • HTML全文
    • 参考文献(31)
    • 相关文章
    • 施引文献
  • 资源附件(0)

/

返回文章
返回