Searching for Pneumothorax in X-Ray Images Using Autoencoded Deep Features

Antonio Sze-To, Abtin Riasatian, Hamid R. TizhoosharXiv 2021 · arXiv

Introduction

Pneumothorax, commonly known as a collapsed lung, occurs when air leaks into the space between the lung and the chest wall. This condition is a medical emergency that requires rapid diagnosis. However, small or subtle collapses are frequently missed on standard chest X-rays, even by experienced radiologists, particularly in high-pressure clinical environments.

While deep learning has shown promise in medical imaging, most existing models function as "black box" classifiers. These systems provide a probability score but lack transparency, making it difficult for clinicians to trust or verify the results. To address this, researchers have developed AutoThorax-Net, a system based on Content-Based Medical Image Retrieval (CBMIR). Instead of simply labeling an image, AutoThorax-Net identifies the most similar historical cases from a massive database, providing radiologists with a "virtual second opinion" backed by tangible evidence.

Methodology: The AutoThorax-Net Pipeline

The AutoThorax-Net framework moves away from traditional end-to-end classification in favor of a multi-stage retrieval process that emphasizes explainability and efficiency.

1. Symmetry Exploitation and Feature Extraction

The system leverages the natural bilateral symmetry of the human chest. Using a pre-trained DenseNet121 model, deep features are extracted from three distinct views: the entire chest X-ray, a cropped image of the left lung, and a cropped/flipped image of the right lung. By aligning the right lung with the left, the model can more effectively compare localized abnormalities. These features are then concatenated into a high-dimensional feature vector (3072-D).

The AutoThorax-Net architecture, illustrating the integration of symmetry-based cropping and the autoencoder for feature compression.

2. Dimensionality Reduction via Autoencoding

To ensure efficient searching across large databases, the 3072-D vector is compressed using an autoencoder. This neural network is trained to reduce the vector to a 256-D "fingerprint" while preserving the most critical diagnostic information. This compression not only speeds up retrieval but also helps filter out noise that might lead to false positives.

3. Similarity Search and Classification

When a new query image is processed, its fingerprint is compared against a database of over 550,000 historical X-rays (sourced from MIMIC-CXR, CheXpert, and ChestX-ray14) using Euclidean distance. The system retrieves the top $k$ most similar images. A final diagnosis is then reached through a majority vote among these retrieved cases, allowing the radiologist to review the similar historical images and their associated reports.

Experimental Results

The researchers evaluated AutoThorax-Net in two scenarios: a semi-automated search (pneumothorax vs. normal) and a fully automated search (pneumothorax vs. all other chest diseases). The retrieval-based approach demonstrated exceptional performance, achieving an Area Under the Curve (AUC) of up to 0.923 in the semi-automated task. Notably, the system outperformed the well-known CheXNet classifier, particularly in its ability to minimize false positives.

ROC curves demonstrating the classification performance of AutoThorax-Net for various values of k, achieving a maximum AUC of 0.923.

Discussion and Conclusion

The success of AutoThorax-Net highlights the power of combining symmetry-aware feature extraction with efficient retrieval. By providing similar historical cases rather than a single score, the system offers a level of explainability that traditional classifiers lack.

One identified challenge is the presence of chest tubes in training data, which can act as a "shortcut" for the AI to identify pneumothorax. Future work will focus on mitigating this bias, exploring alternative distance metrics, and integrating textual radiology reports into the search results to further enhance the clinical utility of this "virtual second opinion."

Summary updated

BibTeX
@misc{szeto2021searching,
  author = {Antonio Sze-To and Abtin Riasatian and Hamid R. Tizhoosh},
  title = {{Searching for Pneumothorax in X-Ray Images Using Autoencoded Deep Features}},
  year = {2021},
  eprint = {2102.06096},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.06096},
}
Figure