Sample-based relationship for assisting diagnosis of pneumonia in medical care

Hongkang Chen, Huijuan Lu, Wenjie Zhu, Ye Zhou, Yudong Yao, Renfeng Wang

Research output: Contribution to journalArticlepeer-review

Abstract

Deep neural networks have succeeded in learning balanced and imbalanced data in the field of pneumonia diagnosis. However, both require separate model designs in their respective domains. The pneumonia recognition methods have poor transfer performance between the two types of image distribution data. This paper proposes BeFormer, which focuses on sample relationships. BeFormer introduces two stages, feature extraction and establishing sample relationships, to solve the problem. Specifically, in the feature extraction stage, a new backbone network called PANet based on Multi-Scale Frequency Domain Attention (FPSA) is proposed. In establishing sample relationships, Deformable Transformer is used in parallel to model the relationships between partial regions of interest among the same and different samples in the pneumonia dataset. These two modules are connected in series to form BeFormer. Without making any changes to the model, this paper conducted experiments using BeFormer on both balanced and imbalanced pneumonia datasets. As compared to traditional pneumonia recognition models, the proposed method achieved significant improvements in the two different data distributions, demonstrating outstanding transfer performance.

Original languageEnglish
JournalMultimedia Tools and Applications
DOIs
StateAccepted/In press - 2024

Keywords

  • Balanced and imbalanced data learning
  • Batch domain learning
  • Pneumonia X-ray
  • Sample relationship
  • Transformer

Fingerprint

Dive into the research topics of 'Sample-based relationship for assisting diagnosis of pneumonia in medical care'. Together they form a unique fingerprint.

Cite this