基于动态卷积的胸部X光片疾病分类算法

Classification Algorithm for Chest X-Ray Diseases Based on Dynamic Convolution

胸部疾病严重威胁人类健康,及时并精准地检测胸部疾病对患者的治疗与康复具有重要意义.胸部疾病经常通过胸部X光片进行诊断,但由于胸部疾病的多样性以及病理特征的复杂性,现有的胸部X光片疾病分类算法存在分类准确度较低、模型复杂度较高等问题.针对以上问题,提出一种基于动态卷积的胸部X光片疾病分类算法.将动态卷积模块加入密集连接网络,在不显著增加网络模型尺寸的前提下,增强网络对多尺度信息的特征提取能力,在提升分类准确度的同时保持高效推理;使用Meta-ACON改进ReLU(rectified linear units)激活函数,通过线性-非线性切换因子自适应地选择是否激活以及使用何种激活函数,从而增强网络的泛化能力;提出加权焦点损失函数,在焦点损失函数的基础上加入权重调整因子,使网络依据分类的难易程度为每种疾病合理分配权重,增大较难分类疾病的损失占比以提高其分类准确度,进而优化整体性能;对数据加载方式进行优化,增大批数据处理量以提升批归一化效果;在测试阶段使用测试时数据增强策略,综合分析多个维度的分类结果,提高分类的准确性与鲁棒性.在chest X-ray14数据集上的实验结果表明,在密集连接网络中加入动态卷积模块、Meta-ACON激活函数、加权焦点损失函数并在实验时优化数据加载方式、使用测试时数据增强策略的算法对14种胸部疾病分类的平均受试者工作特征曲线下面积(area under receiver operating characteristic curve,AUC)值达到0.8361,针对单个疾病标签的AUC值最高可达0.9534,高于目前6种先进算法.实验结果表明,基于动态卷积的胸部X光片疾病分类算法具有分类准确度高、模型鲁棒性强等优势,可良好地适用于胸部X光片疾病分类任务.

Chest diseases are a severe threat to human health.The timely and accurate detection of chest diseases has great significance on patients’treatment and rehabilitation.Chest X-rays are usually used to diagnose chest diseases.However,the existing classification algorithms for chest X-ray diseases are inaccurate due to the diverse chest diseases and their complex pathological characteristics,including low classification accuracy and high model complexity.In response to the above problems,a new classification algorithm for chest X-ray diseases was proposed based on dynamic convolution.The dynamic convolution blocks were added to the densely connected network.The feature extraction capability of the network for multiscale information was enhanced under the premise of not significantly increasing the size of the network model and reducing efficient inference while improving the classification accuracy.The ReLU(rectified linear units)activation function was improved using Meta-ACON,and the generalization ability of the network was enhanced by adaptively selecting the linear-nonlinear switching factor to activate and which activa tion function to use.The weighted focal loss function was proposed based on a weight adjustment factor added into the focus loss function so that the network reasonably allocated the weight for each disease according to the classification difficulty.To improve their classification accuracy,the overall performance was optimized by increasing the percentage loss of difficult-to-classify diseases.To improve the batch normalization,the data loading method was optimized and batch size increased.In the testing phase,a test-time data augmentation was used to analyze the multiple dimensions classification results and increase the accuracy and robustness of the classification.We found that the average value of the area under the receiver operating characteristic curve(AUC)of the proposed algorithm for the 14 chest diseases was 0.8361 after adding dynamic convolution blocks,Meta-ACON activation function,and weighted focus loss function into the densely connected network,optimizing the data loading method and using test-time data enhancement strategy.Furthermore,the AUC value of a single disease label was up to 0.9534,which is higher than that of the existing six state-of-the-art algorithms.Our results are higher than those of the six recent existing advanced algorithms.The experimental results show that the classification algorithms for chest X-ray diseases based on dynamic convolution have a higher classification accuracy and stronger model robustness.Hence,the algorithm could be well applied to the chest X-ray diseases classification task.