并行多尺度特征融合的肺炎CT分割方法

Parallel multiscale feature fusion for pneumonia CT segmentation

针对肺炎CT图像病灶区域不连续以及与正常组织间差异小,导致难以提取边缘特征以及病变区域噪声较多等问题,提出了一种全局与局部并行分割网络。该网络采用分层PVTv2和ResNet50从全局与局部特征角度提取病变区域的特征。并且,在全局与局部网络之间加入双注意力特征融合模块聚合各层特征信息,有效捕捉病变区域的语义信息和空间细节。同时,为提高网络对病变区域噪声的识别能力,还提出了多尺度并行残差解码器,动态调节权重,使网络高效地利用浅层和深层特征。最后,采用边缘监督机制,在具有边缘信息的低层特征上增强病灶边界的感知能力。在COVID-19-CT-Seg公开数据集上进行测试,Dice系数为76.8%、灵敏度为79.4%、特异度为96.4%,相较于Inf-net网络,分别提升了8.6%、10.2%、9.1%。在另一肺炎分割数据集上测试,Dice系数为75.3%。结果表明,对比其他方法,该方法能够有效地分割肺炎CT图像病灶区域,具有较好的泛化性能,同时为进一步肺炎CT量化处理提供方案。

Aiming at the issues of discontinuous lesion regions and minimal differences between pneumonia CT images and normal tissues,which make it difficult to extract edge features and result in more noise in the affected areas,a novel approach of a parallel global and local segmentation network is proposed.This network employs hierarchical PVTv2 and ResNet50 to extract features of the lesion region from both global and local perspectives.Furthermore,a dual-attention feature fusion module is introduced between the global and local networks to aggregate feature information from multiple layers,effectively capturing the semantic details and spatial information of the affected regions.To enhance the network's ability to recognize noise in the lesion areas,a multiscale parallel residual decoder is proposed,which dynamically adjusts weights to efficiently utilize shallow and deep features.Additionally,an edge-supervised mechanism is implemented to enhance the perception of lesion boundaries by leveraging edge information in lower-level features.The proposed method is evaluated on the publicly available COVID-19-CT-Seg dataset,achieving a Dice coefficient of 76.8%,sensitivity of 79.4%,and specificity of 96.4%.These metrics demonstrate respective improvements of 8.6%,10.2%,and 9.1%compared to the Inf-net network.Testing on another pneumonia segmentation dataset yields a Dice coefficient of 75.3%.The results indicate that the proposed method effectively segments the lesion regions in pneumonia CT images,exhibits good generalization performance,provide a solution for further quantification processing of pneumonia CT scans simultaneously.