Anatomy-Guided Weakly-Supervised Abnormality Localization in Chest X-rays.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Creating large-scale annotated medical image datasets for abnormality detection is resource-intensive.
• Weak supervision from radiology reports offers a potential solution but often overlooks anatomical details.
• Existing methods struggle with noisy labels derived from Natural Language Processing (NLP) due to sparsity and ambiguity.

Purpose of the Study:

• To develop an Anatomy-Guided chest X-ray Network (AGXNet) for improved weak annotation in medical image analysis.
• To leverage both pathological observations and anatomy mentions from radiology reports.
• To address label noise and sparsity in NLP-mined weak labels.

Main Methods:

• Proposed a cascaded two-network framework (AGXNet) for identifying anatomical abnormalities and pathological observations.
• Introduced an anatomy-guided attention module to focus the observation network on relevant anatomical regions.
• Employed Positive Unlabeled (PU) learning to handle cases where the absence of a mention does not imply a negative label.

Main Results:

• AGXNet demonstrated effectiveness in localizing diseases and anatomical abnormalities on the MIMIC-CXR dataset.
• Learned feature representations from AGXNet were transferable to the NIH Chest X-ray dataset.
• Achieved state-of-the-art performance in disease classification and competitive results in disease localization.

Conclusions:

• AGXNet successfully addresses limitations of weak annotation in medical imaging by integrating anatomical context.
• The proposed method enhances the utility of radiology reports for training robust anomaly detection models.
• AGXNet shows promise for improving diagnostic accuracy and efficiency in medical image analysis.