Learning to Recognize Thoracic Disease in Chest X-Rays With Knowledge-Guided Deep Zoom Neural Networks

Kun Wang, Xiaohong Zhang, Sheng Huang, Feiyu Chen, Xiangbo Zhang, Luwen HuangfuIEEE Access 2020 · DOI

Introduction

Interpreting Chest X-rays (CXRs) is a critical but challenging task for radiologists, often prone to subjective error and fatigue. Automated deep learning models face three primary hurdles: imaging noise from irrelevant regions (e.g., neck, abdomen), visual similarity between diseased and healthy tissue, and the "small lesion" problem, where crucial features are lost during image downsampling. Traditional models often treat the entire image as a single data point, failing to capture the "coarse-to-fine" diagnostic process used by human experts.

The KGZNet Framework

To address these challenges, the authors propose KGZNet (Knowledge-Guided Deep Zoom Neural Network). This multi-scale architecture mimics a radiologist's workflow by iteratively focusing from a global chest view down to specific lung and lesion regions, guided by medical prior knowledge.

Conceptual workflow of KGZNet, integrating medical knowledge and multi-scale visual information for thoracic disease diagnosis.

Multi-Scale Architecture

KGZNet consists of three specialized branches that process the image at different scales:

  • GlobalNet (Coarse Scale): Processes the entire downsampled CXR to capture global structural context.
  • LungNet (Medium Scale): Uses a modified U-Net to segment lung fields. This acts as a "medical prior," explicitly directing the model to focus on the lungs where thoracic diseases manifest, thereby filtering out background noise.
  • LesionNet (Fine Scale): This "zoom" component uses Class Activation Maps (CAM) from the LungNet to identify the most discriminative (suspicious) areas and crops them at a higher relative resolution.
Detailed architecture of KGZNet, showing (a) the multi-scale zoom view, (b) the extraction of global, lung, and lesion regions, and (c) the MS-KGZ fusion module.

The features from all three branches are concatenated in a FusionNet, allowing the model to make a final diagnosis based on the relationship between global context and local details. Crucially, the system utilizes Weakly-Supervised Learning (WSL), learning to localize lesions using only image-level labels rather than expensive, manually-annotated bounding boxes.

Key Results and Insights

The model was evaluated on the NIH Chest X-ray 14 dataset (112,120 images) and the RSNA Pneumonia Detection Challenge (KRPDC) dataset.

  • State-of-the-Art Performance: KGZNet achieved an average AUC of 0.878 on the NIH dataset, outperforming well-known models like CheXNet (0.841) and AG-CNN (0.871).
  • Superior Localization: Even without bounding box training, the model's "zoomed" regions showed high overlap with ground-truth lesion locations provided by radiologists.
  • Robustness: On the KRPDC dataset, KGZNet improved accuracy by over 3% compared to standard ResNet and DenseNet backbones, demonstrating better handling of "out-of-distribution" data.

Conclusion

KGZNet demonstrates that mimicking human diagnostic intuition through a multi-scale "deep zoom" approach, combined with medical priors, significantly enhances thoracic disease recognition. By effectively filtering noise and preserving fine-grained details of small lesions, it sets a new benchmark for automated CXR analysis. Future work will focus on more sophisticated attentional maps and lightweight versions for clinical deployment.

Summary updated

BibTeX
@article{wang2020learning,
  author = {Kun Wang and Xiaohong Zhang and Sheng Huang and Feiyu Chen and Xiangbo Zhang and Luwen Huangfu},
  title = {{Learning to Recognize Thoracic Disease in Chest X-Rays With Knowledge-Guided Deep Zoom Neural Networks}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3020579},
}
Figure