Automated segmentation and diagnosis of pneumothorax on chest X-rays with fully convolutional multi-scale ScSE-DenseNet: a retrospective study

Qingfeng Wang, Qiyu Liu, Guoting Luo, Zhiqin Liu, Jun Huang, Yuwei Zhou, Ying Zhou, Weiyun Xu, Jie-Zhi ChengBMC Medical Informatics Decis. Mak 2020 · DOI

Introduction

Pneumothorax (PTX) occurs when air enters the pleural space, causing lung collapse. It is a medical emergency requiring rapid intervention. However, identifying PTX on chest X-rays is notoriously difficult; visual evidence often consists of subtle, thin lines easily obscured by anatomical structures like ribs or clavicles. While previous AI efforts focused on image-level diagnosis, they often lacked the precise localization necessary for clinical trust and severity assessment.

Proposed Architecture

The researchers developed MS_scSE_FC-DenseNet, an end-to-end framework designed for pixel-level segmentation and diagnosis. The model builds upon the Fully Convolutional DenseNet (FC-DenseNet) architecture, optimized for parameter efficiency and high-resolution output.

Overview of the MS_scSE_FC-DenseNet architecture and its sub-modules.

The pipeline begins with a multi-scale convolution module that uses parallel kernels (3×3, 5×5, and 7×7) to capture features across different receptive fields. This allows the network to detect both fine pleural lines and larger regions of air accumulation simultaneously.

Attention Mechanisms and Loss Function

A core innovation is the integration of scSE modules (Spatial and Channel Squeeze & Excitation) within every dense block. These modules act as a digital magnifying glass, recalibrating feature maps to emphasize relevant spatial locations and important channels while suppressing background noise from bones and other tissues.

The concurrent spatial and channel squeeze and excitation (scSE) module.

To further improve segmentation of the thin pleural line, the authors introduced a Spatially Weighted Cross-Entropy Loss (SW-CEL). Standard loss functions often fail because the target pixels represent a tiny fraction of the image. SW-CEL assigns higher weights to boundary pixels, forcing the model to prioritize the precision of the lung edge.

Dataset and Training

The study utilized the PX-ray dataset, comprising 11,051 front-view chest X-rays (5,566 PTX cases and 5,485 non-PTX cases). Unlike datasets relying on automated report mining, every PTX image was pixel-wisely annotated by medical students and meticulously verified by an experienced radiologist. The data was split into training (64%), validation (16%), and testing (20%) sets using stratified sampling.

Key Results

The MS_scSE_FC-DenseNet was compared against several state-of-the-art models, including U-Net, SegNet, and DeepLab v3+:

  • Segmentation Performance: The model achieved a Mean Pixel-wise Accuracy (MPA) of 0.93 and a Dice Similarity Coefficient (DSC) of 0.92, significantly outperforming baselines.
  • Diagnostic Accuracy: As a classifier, the system reached 93.45% accuracy with an F1-score of 92.97%.
  • Efficiency: The architecture remains parameter-efficient, achieving superior results with lower computational costs (GFLOPS) than heavier models like DeepLab v3+.

Limitations and Future Work

While the results are promising, the study is retrospective. Future work should include prospective clinical trials to evaluate performance in real-time hospital environments. Additionally, the diagnostic classification currently relies on a fixed pixel-count threshold (50 pixels), which may require further tuning to handle noise across different X-ray hardware. Finally, broader validation across diverse patient demographics and equipment manufacturers is necessary for widespread clinical deployment.

Summary updated

BibTeX
@article{wang2020automated,
  author = {Qingfeng Wang and Qiyu Liu and Guoting Luo and Zhiqin Liu and Jun Huang and Yuwei Zhou and Ying Zhou and Weiyun Xu and Jie-Zhi Cheng},
  title = {{Automated segmentation and diagnosis of pneumothorax on chest X-rays with fully convolutional multi-scale ScSE-DenseNet: a retrospective study}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2020},
  doi = {10.1186/s12911-020-01325-5},
}
Figure