用于胸片分类的自校正特征融合金字塔网络  

Chest X-Ray Classification Based on Self-Calibrated Feature Fusion Pyramid Network

作　　者：宫霄霖 程琦 李锵 Gong Xiaolin;Cheng Qi;Li Qiang(School of Microelectronics,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学微电子学院,天津300072

出　　处：《天津大学学报（自然科学与工程技术版）》2024年第5期511-520,共10页Journal of Tianjin University：Science and Technology

基　　金：国家自然科学基金资助项目(62272337);天津大学自主创新基金资助项目(2021XZC-0024);超声医学工程国家重点实验室开放课题资助项目(2022KFKT004).

摘　　要：胸部疾病的病灶区域面积差异较大,易受健康部位的影响,难以定位,并且用于诊断疾病的X射线胸片样本数据分布不平衡,导致胸部疾病分类网络的分类准确度较低.针对胸部疾病分类任务中仍然存在的问题和挑战,本文提出了一种自校正特征融合金字塔网络.该网络使用自校正卷积增强疾病部位在特征图上以及通道之间的位置联系,在不引入额外参数量的条件下增大了卷积的感受野,避免无病区域的信息干扰;然后通过特征融合金字塔网络整合多尺度图像特征信息,在获取高分辨率特征图的同时,能够准确地定位病变区域,可以更好地识别不同尺度疾病的特征,在多标签分类任务上具有独特的优势;同时引入频率通道注意力机制强化网络对疾病特征的提取能力,在上采样和特征融合变换之前,减少全局平均池化过程中造成的特征丢失;最后提出一种轮次焦点损失函数区分不同种类胸部疾病的训练程度,根据分类难易程度区分样本,并在不同的训练轮次调整权重,以解决胸部疾病样本分布不平衡问题.在ChestX-ray14数据集上的平均AUC值可达0.853,在CheXpert数据集上的平均AUC值可达0.903,超过了近年来较为先进的网络模型.实验结果表明,该网络与传统的胸部疾病分类网络相比能有效地提高胸部疾病的分类精度,并且具有较强的泛化能力.Thoracic diseases are detected in different lesion regions,can affect healthy sites,and are difficult to locate.Furthermore,they exhibit an imbalanced distribution in chest X-ray samples.All of these conditions result in the low accuracy of thoracic disease classification.This paper proposes a self-calibrated feature fusion pyramid network in response to the problems and challenges encountered in thoracic disease classification tasks.This network enhances the positional relationships of disease sites between feature maps and channels with self-calibrated convolutions,improving the convolutional receptive field without introducing additional parameters and avoiding information interference from disease-free regions.Further,the feature fusion pyramid network integrates multiscale features to locate the lesion regions accurately and acquire high-resolution feature maps,enabling better recognition of disease features at different scales and leading to unique advantages in multilabel classification tasks.Moreover,the introduction of frequency channel attention reinforces the capability of the network to extract disease features simultaneously before upsampling and feature fusion,which reduces the loss of features during global average pooling.Finally,an epoch focal loss with different degrees of training across thoracic diseases is proposed to address the sample imbalance,which can distinguish the samples according to the classification difficulty and adjust the weights in different training epochs.The average area under curve(AUC)value can reach 0.853 on the ChestX-ray14 dataset and 0.903 on the CheXpert dataset,outperforming the most advanced network models in recent years.The experimental results show that the network can effectively improve the classification accuracy of thoracic diseases and has a stronger generalization capability than state-of-the-art classification networks.

关 键 词：胸部疾病 自校正卷积 特征融合金字塔网络 频率通道注意力 轮次焦点损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]