Weakly supervised pneumonia localization in chest X-rays using generative adversarial networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pneumonia localization on chest X-rays (CXRs) is challenging due to visual ambiguity and complex anatomy.
• Existing methods often require expert annotations or lack spatial information preservation.
• Accurate localization is crucial for timely diagnosis and as an aid to radiologists.

Purpose of the Study:

• To develop a weakly supervised, generative adversarial network (GAN)-based approach for automatic pneumonia localization on CXRs.
• To generate pixel-wise abnormality maps, highlighting pneumonia regions without bounding boxes.
• To overcome limitations of existing methods by not requiring location annotations.

Main Methods:

• Utilized the Wasserstein GAN framework to generate pseudo-normal CXRs from abnormal ones.
• Trained on an unpaired dataset of normal and abnormal CXRs, requiring no location annotations.
• Incorporated prior knowledge and constraints to enhance localization performance.

Main Results:

• Achieved an ROC area under the curve (AUC) of 0.77 with the best performing method.
• Demonstrated an 85% abnormality detection rate.
• Visual results showed effective highlighting of abnormal regions across various scenarios.

Conclusions:

• Presented a novel GAN-based method for pneumonia localization on CXRs.
• The approach is weakly supervised, eliminating the need for expensive location annotations.
• The method effectively produces pixel-level abnormality maps, proving its efficacy through quantitative and qualitative results.