多视角胸部X光图像的数据预处理与分类研究
A Systematic Preprocessing Pipeline and Data Preprocessing and Classification Study for Multi-View Chest X-Rays
王冠徽 1苏放1
作者信息
- 1. 北京邮电大学人工智能学院，北京 100876
- 折叠
摘要
多视角胸部X光图像在辅助诊断中价值巨大，但真实数据集（如MIMIC-CXR）常存在罕见视图的长尾分布、标签冲突与缺失等严重噪声，制约了特征的有效融合。为此，本文提出一套系统化的视图标签预处理流程。基于高效的辅助视图分类器，实现了罕见视角的相似性映射、冲突标签纠正及缺失标签补全。为评估预处理的价值，本文对比了单视角、多视角拼接与加权融合三种基线模型。实验表明，多视角基线稳定优于单视角；在预处理之后的高质量数据上，后两类基础多视角融合模型的有一定的性能提升。本文工作为多视角诊断提供了可靠基线，并提供了较高质量的数据基础。
Abstract
Multi-view chest X-ray (CXR) images hold significant value in computer-aided diagnosis. However, real-world datasets (e.g., MIMIC-CXR) frequently suffer from severe noise, including long-tail distributions of rare views, label conflicts, and missing annotations, which hinders effective feature fusion. To address these challenges, this paper proposes a systematic view label preprocessing pipeline. Leveraging an efficient auxiliary view classifier, we implement similarity mapping for rare views, correction of conflicting labels, and completion of missing annotations. To evaluate the efficacy of this preprocessing, we compare three baseline models: single-view, multi-view concatenation, and weighted fusion. Experimental results demonstrate that multi-view baselines consistently outperform the single-view approach. Furthermore, the basic multi-view fusion models exhibit measurable performance gains when trained on the high-quality data produced by our preprocessing. This work establishes a reliable baseline for multi-view CXR diagnosis and provides a robust data foundation for future research.关键词
深度学习/胸部X光/多视角融合/数据预处理/Key words
Deep Learning/Chest X-Ray/Multi-View Fusion/Data Preprocessing/引用本文复制引用
王冠徽,苏放.多视角胸部X光图像的数据预处理与分类研究[EB/OL].(2026-04-21)[2026-04-22].http://www.paper.edu.cn/releasepaper/content/202604-156.学科分类
医学研究方法/计算技术、计算机技术
评论