【作者】 吴静；

【导师】 张子曙；

【作者基本信息】 中南大学 ， 影像医学与核医学， 2022， 博士

【摘要】 背景与目的:在18-氟-氟脱氧葡萄糖正电子发射断层扫描/计算机断层扫描(18-fluorine-fluorodeoxyglucose positron emission tomography/computed tomography scan,18F-FDG PET/CT)上准确区分恶性和良性肺结节,对于分流不确定的肺结节患者非常重要。机器学习一直是评估肺结节的有力工具,但研究主要集中于肺部CT扫描上,且旨在预后的研究大多仍依靠手工特征,而不是深度学习。研究1评估了卷积神经网络(Convolutional neural network,CNN)利用18F-FDG PET/CT图像区分恶性和良性肺结节的有效性。研究2探讨了使用机器学习分析治疗前的18F-FDG PET/CT图像,以预测肺癌的进展和总体生存率(overall survival,OS)的价值。材料与方法:研究1:回顾性研究3个机构中因不确定肺结节接受18F-FDG PET/CT检查的的患者。679名患者符合纳入标准,包含868个病变,其中602个结节为恶性,262个为良性。对他们的18F-FDG PET/CT图像进行了手工分割。病变以7:2:1的比例分成训练集、验证集和测试集,并使用CT和PET输入训练一个包含Efficient Net B4架构的CNN的集成模型。模型在测试集上的表现根据接收者操作曲线下面积(Area under the receiver operating curve,AUC)、准确度、敏感度和特异度进行评估,并与五位专家人工阅片诊断的结果进行比较。研究2:回顾性研究3个机构中有治疗前18F-FDG PET/CT检查的肺癌患者。对病变进行手动和自动分割,并使用18F-FDG PET/CT输入训练CNN以预测肺癌的进展。使用AUC、准确度、敏感度和特异度来评估模型性能。从CNN中提取图像特征,并通过传统影像组学提取特征,构建随机生存森林模型(Random survival forests,RSF)来预测OS。使用一致性指数(Concordance index,C-index)和Brier综合评分(Integrated Brier score,IBS)来评估模型对OS的预测。结果:研究1:18F-FDG PET/CT集成模型的测试AUC为0.81,准确度为0.74【95%置信区间(Confidence interval,CI):0.64,0.82】,敏感度为0.72(95%CI:0.60,0.82),特异度为0.78(95%CI:0.59,0.90)。该模型在准确度(0.74 vs.0.59,0.58;P=0.04,0.03)和敏感度(0.72vs.0.51,0.56;P=0.01,0.04)方面优于五位专家中的两位。研究2:共纳入1168个肺癌肿瘤(965名患者)。其中792个肿瘤有进展,376个无进展。对于预测进展风险,使用手动分割的PET+CT集成模型(准确度为0.790,AUC为0.876)与CT模型(准确度为0.723,AUC为0.888)表现相似,与PET模型(准确度为0.664,AUC为0.669)相比更优。对于使用深度学习特征的OS预测,PET+CT+临床的RSF集成模型(C-index=0.737)的表现与CT模型(C-index=0.730)相似,优于PET(C-index=0.595)和临床(C-index=0.595)模型。用传统影像组学特征构建的RSF模型与用深度学习特征构建的RSF模型性能相当。结论:使用18F-FDG PET/CT数据训练的CNN在区分肺结节良恶性方面表现良好,与专家的性能相当或更优。采用治疗前18F-FDG PET/CT数据训练的CNN在预测肺癌的进展和OS方面表现良好。图10幅,表11个,参考文献129篇更多还原

【Abstract】 Background and Purpose:Accurate differentiation of malignant from benign pulmonary nodules on 18-fluorine-fluorodeoxyglucose positron emission tomography/computed tomography scan(18F-FDG PET/CT)is crucial for triaging patients for more extensive work-up or less aggressive management of an indeterminate lung nodule.Machine learning has been a powerful tool in evaluating lung nodules but has focused primarily on lung CT scans,and those aimed at prognostication have mostly relied on handcrafted features rather than deep learning.Study 1 evaluated the effectiveness of convolutional neural network(CNN)in distinguishing malignant from benign lesions using images from 18F-FDG PET/CT scans.Study 2 explored the value of using machine learning to analyze pre-treatment 18F-FDG PET/CT scans to predict lung cancer progression and overall survival(OS).Materials and methods:Study 1:A retrospective review was conducted across three institutions identifying patients that had received a 18F-FDG PET/CT as a part of pre-procedure work-up of an indeterminate lung nodule.679 patients met the inclusion criteria and 868 lesions,602 malignant and 262 benign,were manually segmented from their 18F-FDG PET/CT scans.Lesions were split 7:2:1 between training,validation,and test sets,and an ensemble model containing convolutional neural nets with Efficient Net B4 architectures was trained using CT and PET inputs.Model performance on the test set was evaluated on area under the receiver operating curve(AUC),accuracy,sensitivity,and specificity and compared to the performance of five experts.Study 2:A retrospective review across three institutions identified patients who had a pre-procedure 18F-FDG PET/CT and an associated lung cancer diagnosis.Lesions were manually and automatically segmented,and CNNs were trained using 18F-FDG PET/CT inputs to predict cancer progression.Performance was evaluated using AUC,accuracy,sensitivity,and specificity.Image features were extracted from CNNs and by radiomics feature extraction,and random survival forests(RSF)were constructed to predict OS.Concordance index(C-index)and Integrated Brier score(IBS)were used to evaluate OS prediction.Results:Study 1:The PET/CT ensemble model achieved a test AUC of 0.81,accuracy of 0.74 [95% confidence interval(CI): 0.64,0.82],sensitivity of 0.72(95% CI: 0.60,0.82),and specificity of 0.78(95% CI: 0.59,0.90),PPV of0.88,and NPV of 0.55.The model outperformed two out of five experts on accuracy(0.74 vs.0.59,0.58;P = 0.04,0.03)and sensitivity(0.72 vs.0.51,0.56;P = 0.01,0.04).Study 2:1168 tumors(965 patients)were identified.792 tumors had progression and 376 were progression-free.The most common subtypes were adenocarcinoma(n = 740)and squamous cell carcinoma(n = 179).For progression risk,the PET + CT ensemble model with manual segmentation(accuracy = 0.790,AUC = 0.876)performed similarly to the CT only(accuracy = 0.723,AUC = 0.888)and better compared to the PET only(accuracy = 0.664,AUC = 0.669)model.For OS prediction with deep learning features,the PET + CT + clinical RSF ensemble model(C-index = 0.737)performed similarly to the CT only(C-index =0.730)and better than the PET only(C-index = 0.595),and clinical only(C-index = 0.595)models.RSF models constructed with radiomics features had comparable performance to those with deep learning features.Conclusion:CNNs trained using 18F-FDG PET/CT data can achieve good performance in differentiation of malignant from benign lung nodules that is comparable or better than expert performance.These models have the potential to improve clinician’s ability to triage and diagnose patients.CNNs trained using pre-treatment 18F-FDG PET/CT performed well in predicting lung cancer progression and OS.The prognostic models could inform treatment options and improve patient care.更多还原