基于双侧学习分支聚合上下文和局部特征的肝包虫病病灶分割

Segmentation on Pathological Regions of Hepatic Echinococcosis Based on Bilateral Learning Branch Aggregation of Context and Local Features

摘要: 利用人工智能算法对医学CT图像中的肝包虫病变区域进行检测和分析，对医生的工作效率提高具有重要意义。然而，由于数据集的大小有限、局部细节模糊、形状不规则等因素，CT图像中肝包虫病区域的分割仍然具有挑战性。因此，本文提出了一种新的肝包虫病分割网络，该网络结合了两个特征提取分支以实现速度和准确性的平衡。首先我们设计了一个上下文分支(CB)，通过类似Transformer的模块来保留尺度不变的全局上下文信息。然后，基于深度聚合(DAP)模块的浅层细节分支(DB)来提供详细信息。我们在私有数据集上进行了大量实验。实验结果表明，所提出的网络优于现有方法，并在准确性和推理速度之间取得了良好的平衡。

Abstract: Utilizing artificial intelligence algorithms for detecting and analyzing hepatic echinococcosis regions in medical CT images holds significant importance in improving the efficiency of medical practitioners. However, segmentation of echinococcosis regions in CT images remains challenging due to factors such as limited dataset size, blurred local details, and irregular shapes. Therefore, this paper proposes a novel echinococcosis lesion segmentation network that integrates two feature extraction branches to achieve a balance between speed and accuracy. Firstly, we design a contextual branch (CB) to preserve scale-invariant global contextual information using a Transformer-like module. Additionally, a shallow detail branch (DB) based on the Deep Aggregation Pyramid (DAP) module is introduced to provide detailed information. Extensive experiments are conducted on a private dataset. Experimental results demonstrate that the proposed network outperforms existing methods and achieves a good balance between accuracy and inference speed.