基于支气管内超声多模态影像的胸内良恶性淋巴结预测模型研究

There is a wide spectrum of benign and malignant disease in intrathoracic lymph nodes. The treatment strategy and prognosis are affected by the diagnosis. Endobronchial ultrasound (EBUS) multi-modal imaging, including gray scale ultrasound, blood flow Doppler and elastography, is a noninvasive method in the diagnosis of intrathoracic lymph nodes and can supplement the disadvantages of pathology results of EBUS-transbronchial needle aspiration (EBUS-TBNA), which has significant application value in the clinical practice. The project intends to investigate the effect of cardiopulmonary exercise on EBUS multi-modal imaging by monitoring the signal of breath and electrocardio simultaneously, then directing the selection of representative image from dynamic videos with the assistance of computer. The representative image selected by the computer will be compared with both the pathological features of resected lymph node and the image selected by experts to investigate the concordance. A diagnostic evaluation system of EBUS multi-modal imaging for intrathoracic benign and malignant lymph nodes will be established through the semi-quantitative and quantitative analysis of the representative images selected by the computer. EBUS multi-modal imaging database containing 6,000 benign and malignant lymph nodes will be constructed to train deep learning network for automatic extraction and diagnosis of region of interest. A predictive model of intrathoracic benign and malignant lymph nodes will be established and EBUS multi-modal imaging data of 1,000 lymph nodes will be performed to validate the model. Therefore, this study can provide a noninvasive method in diagnosing intrathoracic benign and malignant lymph nodes with a reliable theoretical basis based on EBUS multi-modal imaging to optimize the mode of diagnosis and treatment.

胸内良恶性淋巴结疾病谱广泛，其诊断结果影响疾病治疗策略和预后判断。支气管内超声（EBUS）多模态影像特征（灰阶声像、血流多普勒、弹性成像）可对胸内良恶性淋巴结进行无创诊断，补充EBUS引导下经支气管针吸活检病理诊断结果不足，具有重要临床应用价值。本项目拟通过同步监测呼吸和心电信号探讨心肺运动对EBUS多模态影像的影响，指导计算机动态视频代表性图像选择，并与手术切除淋巴结病理特征及专家选择代表性图像进行一致性验证；通过对机选代表性图像半定量和定量分析，建立和验证EBUS多模态影像对胸内良恶性淋巴结诊断评价体系。通过构建6000例胸内良恶性淋巴结EBUS多模态影像数据库，训练深度学习网络用于目标区域自动提取和诊断，建立胸内良恶性淋巴结预测模型，并对1000例淋巴结EBUS多模态影像数据进行验证。因此，本研究可为EBUS多模态影像无创诊断胸内良恶性淋巴结提供可靠理论依据，优化诊疗模式。

胸内良恶性淋巴结疾病谱广泛，其诊断结果影响疾病的治疗策略和预后判断。支气管内超声（EBUS）多模态影像可结合不同超声成像模式（灰阶声像、血流多普勒、弹性成像）对胸内良恶性淋巴结进行无创诊断，具有重要的临床应用价值。本项目主要完成下述工作：①回顾性分析胸内良恶性淋巴结的EBUS多模态影像特征，建立了鉴别胸内淋巴结良恶性质的定性评价体系，并进行前瞻性验证，明确了EBUS多模态影像在评估胸内良恶性淋巴结的诊断价值。②通过对EBUS多模态影像视频特征的机器学习分析，实现了计算机辅助提取EBUS视频中的代表性图像，其一致性达到专家选图水平。③构建了胸内良恶性淋巴结EBUS多模态影像标注库，利用深度学习技术训练EBUS多模态影像诊断神经网络，分别完成了基于图片和视频的EBUS多模态影像辅助诊断模型，实现专家级水平的自动识别和准确诊断。因此，通过本研究可为EBUS多模态影像无创诊断胸内良恶性淋巴结和指导经支气管针吸活检提供理论依据和技术可行性，优化胸内良恶性淋巴结疾病诊疗模式。