融合边缘校准与混合注意力的儿童肺炎CT图像分割网络

Pediatric Pneumonia CT Image Segmentation Network with Edge Calibration and Hybrid Attention

摘要:

肺炎是威胁儿童健康的主要疾病之一，精准的肺炎计算机断层扫描(Computed Tomography, CT)图像分割对儿童肺炎的早期诊断和治疗至关重要。但由于儿童肺炎病灶破碎分散以及边界高度不规则，现有方法在儿童肺炎分割中表现不佳。为解决上述问题，本文提出了一种融合边缘校准与混合注意力的儿童肺炎CT图像分割网络。该网络基于U型架构，结合边缘校准和多尺度混合注意力模块，能增强对局部边界和全局零散病灶的学习。同时，采用尺寸感知选择性融合模块，自适应地融合不同层次的特征，改善小病灶的分割效果。实验结果表明，该方法在临床儿童肺炎CT数据集上的表现优于其他方法，且在两个公开新冠肺炎CT数据集上性能良好，验证了其泛化能力。

Abstract:

Pneumonia is one of the major diseases that threaten children’s health. Accurate pneumonia of Computed Tomography (CT) image segmentation is critical for the early diagnosis and treatment of pneumonia in children. However, due to the fragmentation and dispersion of pediatric pneumonia lesions along with the highly irregular boundaries, the existing methods underperform in the pediatric pneumonia segmentation. To address these issues, a pediatric pneumonia CT image segmentation network with edge calibration and hybrid attention is proposed in this paper. Based on the U-shaped architecture, the network integrates an edge calibration module with a multi-scale hybrid attention module to enhance the learning of both local boundaries and global scattered lesions. Simultaneously, a dimension-aware selective integration module is introduced to adaptively fuse features of different levels, thereby improving the segmentation effect of small lesions. The experimental results demonstrate that the proposed network outperforms other methods on the clinical pediatric pneumonia CT dataset, and has good performance on two public COVID-19 CT datasets, which verifies its generalization ability.