基于ConvNeXt模型的胸部X线图像的疾病分类与可视化

Disease classification and visualization on chest X-ray images based on the ConvNeXt model

目的 胸部X线是临床实践中常见的胸部疾病筛查和诊断方式。由于放射科医生长时间阅片容易视觉疲劳以及医疗资源分配不均衡的问题,导致误诊和漏诊的情况时有发生。针对这一问题,本研究运用深度学习技术,提出了一个基于ConvNeXt模型的胸部X线图像的疾病检测方法,旨在提高胸部疾病诊断准确度、减轻误诊风险并提高医生工作效率。方法 利用大规模公开胸部X线图像数据集ChestX-ray14训练ConvNeXt模型,该模型在ResNet模型的基础上,融合了视觉Transformer结构的优势,可以有效提高模型的特征提取和识别能力,同时以AUC(ROC曲线下方的面积)作为模型性能的评价指标,与已有的分类模型CheXNet、ResNet及Swin Transformer进行了对比。此外,通过引入Grad-CAM可视化方法,利用卷积神经网络特征图的梯度信息生成胸部X线图像的类激活热力图,实现对病灶区域的定位,从而提高医生的诊断效率。结果 基于ConvNeXt模型的诊断方法在识别14种胸部疾病时平均AUC值可达0.842,特别在识别积液(AUC值为0.883)、水肿(AUC值为0.902)和疝气(AUC值为0.942)等疾病时表现较为令人满意。结论 本文提出的方法在胸部X线图像的疾病检测中具有较好的性能,是一种对胸部X线图像进行胸部疾病诊断进而协助医生提高工作效率的有益尝试。

Objective Chest X-ray imaging is a prevalent method for screening and diagnosing chest diseases in clinical practice.However,the likelihood of misdiagnosis and missed diagnosis is exacerbated by radiologists' fatigue from prolonged image reading and uneven distribution of healthcare resources.To address this issue,this study employs deep learning techniques to develop a chest disease detection method based on the ConvNeXt model,aiming to enhance diagnostic accuracy,mitigate the risk of misdiagnosis,and improve physician efficiency.Methods This research utilizes the large-scale public dataset ChestX-ray14 to train a ConvNeXt model,which builds upon the ResNet model and integrates the advantages of visual Transformer structures.This enhancement enables more effective feature extraction and recognition capabilities.The model's performance is evaluated by using the AUC(area under the receiver operating characteristic curve),comparing it with existing classification models such as CheXNet,ResNet,and Swin Transformer.Furthermore,the study incorporates Grad-CAM visualization to generate heatmap-based class activation maps from convolutional neural network feature gradients.These heatmaps facilitate the localization of diseased regions in chest X-rays,thereby enhancing diagnostic efficiency.Results The proposed ConvNeXt-based diagnostic approach achieves an average AUC value of 0.842 when identifying 14 types of chest diseases.It demonstrates particularly promising performance in detecting conditions like pleural effusion(AUC value of 0.883),edema(AUC value of 0.902),and hernia(AUC value of 0.942).Conclusions The proposed method exhibits strong performance in chest disease detection from chest X-rays,serving as a beneficial attempt to assist physicians in improving their diagnostic efficiency through AI-assisted chest X-ray analysis.