Contralaterally Enhanced Networks for Thoracic Disease Detection

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Chest X-ray interpretation is challenging due to low contrast and overlapping tissues.
• Similar anatomical structures exist in the left and right chest, offering potential for diagnostic insights.
• Existing disease detection methods can be improved by incorporating contextual information.

Purpose of the Study:

• To develop a deep end-to-end module for enhancing disease detection in chest X-rays by exploiting contralateral context.
• To improve the feature representations of disease proposals using information from the opposite chest side.
• To create a method integrable with both fully and weakly supervised disease detection frameworks.

Main Methods:

• Utilized a spatial transformer network guided by the spine line to extract contralateral patches.
• Developed a fusion module using additive and subtractive operations to combine features from disease proposals and contralateral patches.
• Integrated the module into existing disease detection frameworks for evaluation.

Main Results:

• Achieved 33.17 AP50 on a private chest X-ray dataset (31,000 images).
• Demonstrated state-of-the-art performance in weakly-supervised disease localization on the NIH chest X-ray dataset.
• The contralateral context module significantly enhanced feature representations for disease detection.

Conclusions:

• The proposed deep learning module effectively utilizes contralateral context for improved chest X-ray disease detection and localization.
• This approach offers a significant advancement for both fully and weakly supervised diagnostic frameworks.
• The method shows promise for enhancing the accuracy and reliability of automated radiological assessments.