Automatic Localization and Identification of Thoracic Diseases from Chest X-rays with Deep Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning for chest X-ray (CXR) analysis faces challenges with limited labeled data for disease locations and poor model transferability.
• Existing datasets often require extensive, error-prone annotation of specific disease locations.

Purpose of the Study:

• To address the limitations of data scarcity and model transferability in deep learning-based CXR disease detection.
• To develop a robust two-stage deep learning model for accurate disease localization and identification in CXRs.

Main Methods:

• A novel bounding box dataset was created, marking anomalous regions without specific disease labels to reduce annotation effort and errors.
• A two-stage deep learning model integrating region proposal network, feature pyramid network, and multi-instance learning was developed.
• The model was trained and validated on the new dataset and CheXpert, then tested on the ChestX-ray14 dataset.

Main Results:

• The model achieved a mean AUC of 0.912 on the CheXpert validation set, outperforming the baseline by 0.021.
• On an external test set, the model achieved a mean AUC of 0.784, surpassing the state-of-the-art model's 0.773.
• Disease localization performance was comparable to existing state-of-the-art models.

Conclusions:

• The developed model demonstrates strong transferability across different CXR datasets.
• The novel bounding box dataset design offers an effective alternative for creating disease localization datasets, reducing annotation burden.