Lung Segmentation Reconstruction Based Data Augmentation Approach for Abnormal Chest X-ray Images Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models for chest X-ray analysis require extensive data for lesion prediction.
• Current data augmentation methods for medical imaging are limited, hindering model performance.

Purpose of the Study:

• To propose a novel data augmentation technique for chest X-ray images using a Lung Segmentation Reconstruction (LSR) module.
• To improve the diagnostic accuracy of deep learning models for cardiopulmonary diseases.

Main Methods:

• Generated synthetic healthy chest X-ray images using an abnormal image as a reference via the LSR module.
• Developed a data augmentation strategy by combining whole images, lung regions, and abnormal regions.
• Integrated enhanced data into a classification model for disease prediction.

Main Results:

• The proposed abnormality enhancement method improved baseline model performance on detecting consolidation and pleural effusion.
• Demonstrated the utility of healthy chest X-ray images within datasets for improving diagnostic predictions.
• Showcased the effectiveness of combining different image regions for enhanced prediction accuracy.

Conclusions:

• The LSR module offers a valuable approach for data augmentation in chest X-ray analysis.
• Integrating healthy and abnormal image data regions can significantly boost deep learning model performance.
• This technique holds potential for improving the diagnosis of various cardiopulmonary conditions.