Learning to detect chest radiographs containing pulmonary lesions using visual attention networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Automated lung nodule detection on chest radiographs shows promise but requires extensive manually annotated data.
• Picture Archiving and Communication Systems (PACS) provide access to large volumes of clinical imaging data.
• Natural language processing can extract binary labels for pulmonary lesions from clinical reports at scale.

Purpose of the Study:

• To propose two novel neural network architectures for detecting pulmonary lesions in chest radiographs.
• To leverage large datasets of weakly-labeled images combined with a smaller set of manually annotated radiographs.
• To improve lesion localization performance using visual attention feedback during training.

Main Methods:

• Developed two novel neural network architectures for chest radiograph analysis.
• Utilized a combination of weakly-labeled and manually annotated chest radiographs for training.
• Implemented a saliency map approach and a recurrent attention model with reinforcement learning for lesion detection.

Main Results:

• The proposed methods were evaluated on a repository of over 430,000 historical chest radiographs.
• Performance was compared against architectures using only weakly-labeled or only annotated images.
• Both novel architectures demonstrated effective lesion detection capabilities.

Conclusions:

• The developed neural networks show potential for automated pulmonary lesion detection in chest radiographs.
• Combining weakly-labeled and annotated data offers a viable strategy for training robust detection models.
• The proposed methods advance the field of AI-assisted radiological diagnosis.