【作者】 王斌；

【导师】 胡凌燕； 陈中书；

【作者基本信息】 南昌大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 结核病是全球范围内导致死亡的主要传染病之一,发病率高的地区主要分布在欠发达国家。支气管内膜结核(Endobronchial Tuberculosis,EBTB)是结核病的一种特殊类型,由于早期症状隐匿,容易漏诊和误诊,导致患者错过最佳治疗时机。随着人工智能在智能诊断领域的发展,EBTB的人工智能辅助诊断逐渐受到关注,但目前高效的人工智能辅助诊断方法仍较少。针对这一问题,本文在原有数据集的基础上进一步完善了EBTB医学图像数据集,并提出了ResNet34-MSCBAM网络模型。在模型训练过程中,结合医学临床诊断经验构建了一种复合损失函数,引导模型更加精准地关注关键病灶区域的颜色特征,从而提升分类性能。基于训练完成的模型,本文开发了EBTB智能辅助诊断系统及网络诊断平台,实现了支气管镜下EBTB的实时智能辅助诊断。首先,本文完善了高质量EBTB医学图像数据集。根据支气管镜下EBTB图像采集标准,整理并标注了来自江西省胸科医院和景德镇第五人民医院的24996张EBTB医学图像,其中包括21092张正常型、669张炎症浸润型(Ⅰ型)、2211张溃疡坏死型(Ⅱ型)、637张肉芽增殖型(Ⅲ型)、148张瘢痕狭窄型(Ⅳ型)、0张管壁软化型(Ⅴ型)和239张淋巴结痿型(Ⅵ型)。每一张EBTB图像均由两名工作10年以上的内镜医生进行标注。最终,经过严格的数据筛选与增强处理,构建了包含19774张高质量EBTB医学图像的数据集。其次,为提升对EBTB图像深层次特征的捕捉能力,本文对CBAM模块进行优化并提出了ResNet34-MSCBAM模型。该模型创新地在ResNet34基础上引入多尺度通道和空间注意力机制,有效融合不同尺度的特征信息,增强了模型对图像细节和全局信息的捕获能力。实验结果表明,该模型在EBTB测试集上的整体分类准确率达到0.864,Ⅰ型图像的诊断敏感度达到0.829,单张EBTB图像诊断时间达到0.049秒。再次,为了提升ResNet34-MSCBAM模型在Ⅰ型EBTB图像分类中的敏感性,本文在分析医学专家如何识别诊断Ⅰ型图像的基础上,将临床医学专家诊断Ⅰ型EBTB图像的医学专业知识融入到损失函数设计中。在设计损失函数时,一方面,针对Ⅰ型EBTB图像的关键病理特征——I型EBTB图像带有红肿现象,提出了红色特征损失函数。该损失函数通过衡量模型提取的特征图与输入图像红色通道之间的差异,强化了模型在训练过程中对红色通道的关注,从而提升了模型对Ⅰ型EBTB相关病灶的识别能力。另一方面,考虑到数据集类别不平衡问题,设计了加权交叉熵损失函数,通过为不同类别分配权重,减轻样本数量不均衡对模型训练的影响。最后将这两个损失函数通过可训练的权重参数进行加权组合,构建了一种复合损失函数。该复合损失函数在训练过程中帮助模型聚焦关键病灶区域,并充分结合自动提取的特征。实验结果表明,该方法使得模型在EBTB测试集上的整体分类准确率提高至0.872,Ⅰ型图像的诊断敏感度提升至0.846。最后,基于ResNet34-MSCBAM网络模型,本文开发了EBTB智能辅助诊断系统和EBTB网络诊断平台。EBTB智能辅助诊断系统的主要功能是在医生进行支气管镜检查与治疗过程中,对实时采集的疑似病灶图像进行自动诊断与分类,结果即时呈现在系统图形用户界面。该系统的硬件配置包括双USB脚踏开关、高清视频采集卡和PC机,系统软件图形用户界面采用Python语言开发并基于Py Qt5框架实现,旨在提升医生的诊断效率并减轻工作负担。EBTB网络诊断平台主要提供远程在线诊断服务,诊断模型部署于云端,用户可以通过网站上传EBTB图片进行自动分析,诊断结果实时生成并展示,帮助欠发达地区的患者获得准确诊断,从而有效减少漏诊与误诊的发生。更多还原

【Abstract】 Tuberculosis is one of the leading infectious diseases causing death globally,with the highest incidence rates concentrated in underdeveloped countries.Endobronchial tuberculosis(EBTB)is a particular form of tuberculosis that is often missed or misdiagnosed due to its subtle early symptoms,causing patients to miss the best time for treatment.With the development of artificial intelligence(AI)in intelligent diagnosis,AI-assisted diagnosis of EBTB has gradually gained attention.However,efficient AI-assisted diagnostic methods for EBTB are relatively few.To address this issue,this study further refines the EBTB medical image dataset and proposes the ResNet34-MSCBAM model.During the model training process,a composite loss function was constructed by combining medical clinical diagnostic experience to guide the model in focusing more precisely on the color features of key lesion areas,thereby improving classification performance.Based on the trained model,this study developed the EBTB intelligent auxiliary diagnosis system and the online diagnostic platform,achieving real-time intelligent auxiliary diagnosis of EBTB under bronchoscopy.First,this study enhanced the quality of the EBTB medical image dataset.According to the collection standards of EBTB images under bronchoscopy,this study sorted and annotated 24996 EBTB medical images from Jiangxi Chest Hospital and Jingdezhen Fifth People’s Hospital,including 21092 normal type images,669 images of the inflammatory infiltration type(Type I),2211 of the ulcerative necrotic type(Type II),637 of the granulomatous proliferative type(Type III),148 of the cicatricial stenosis type(Type IV),0 of the bronchial wall softening type(Type V),and 239 of the lymphatic fistula type(Type VI).Each EBTB image was annotated by two endoscopists with over 10 years of experience.After rigorous data filtering and enhancement processing,a high-quality dataset containing 19774 EBTB medical images was constructed.Second,in order to enhance the ability to capture deep level features from EBTB images,this study optimized the CBAM module and proposed the ResNet34-MSCBAM model.This model innovatively introduced multi-scale channel and the spatial attention mechanism based on ResNet34,effectively integrating feature information of different scales and enhancing the model’s ability to capture both image details and overall information.The experimental results demonstrated that the model achieved an overall classification accuracy of 0.864 on the EBTB test set,with a diagnostic sensitivity of 0.829 for Type I images,and a diagnostic time per EBTB image of 0.049 seconds.Third,in order to improve the sensitivity of the ResNet34-MSCBAM model in classifying Type I EBTB images,this study incorporated clinical expert knowledge into the design of the loss function by analyzing how clinical experts diagnose Type I images.In designing the loss function,two aspects were considered: On the one hand,based on the critical pathological feature of Type I EBTB images—characterized by redness and swelling—a red feature loss function was proposed.This loss function strengthened the model’s focus on the red channel during the training process by measuring the differences between the red channels of the feature image extracted by the model and the input image,thereby enhancing the model’s ability to identify lesions related to Type I EBTB.On the other hand,to address the class imbalance problem in the dataset,a weighted cross-entropy loss function was designed,assigning different weights to various classes to mitigate the impact of sample imbalance on model training.The two loss functions were combined using a trainable weighting parameter,forming a composite loss function.This composite loss function helped the model focus on key lesion areas while leveraging automatically extracted features.Experimental results showed that this approach improved the overall classification accuracy of the model to0.872 on the EBTB test set,with the sensitivity for Type I images increasing to 0.846.Finally,based on the ResNet34-MSCBAM model,this study developed the EBTB intelligent auxiliary diagnosis system and the EBTB online diagnostic platform.The primary function of the EBTB intelligent auxiliary diagnosis system is to provide automatic diagnosis and classification of real-time lesion images collected during bronchoscopy examination and treatment,with results displayed instantly on the system graphical user interface.The hardware configuration of the system includes a dual USB foot pedal,a high-definition video capture card,and a PC.The system’s graphical user interface is developed in Python and implemented with the Py Qt5 framework,aiming to improve doctors’ diagnostic efficiency and reduce their workload.The EBTB online diagnostic platform primarily provides remote online diagnosis services.The diagnostic model is deployed on the cloud,allowing users to upload EBTB images via the website for automatic analysis,with diagnostic results generated and displayed in real time.This platform helps patients in underdeveloped regions receive accurate diagnoses,thus effectively reducing the occurrence of missed and misdiagnoses.更多还原