Multi-task weak supervision enables anatomically-resolved abnormality detection in whole-body FDG-PET/CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Supervised machine learning for whole-body FDG-PET/CT analysis is hindered by limited labeled data and large exam sizes.
• Existing methods struggle with the granularity required for comprehensive clinical decision support.

Purpose of the Study:

• To develop a weak supervision framework for extracting granular regional abnormality labels from free-text radiology reports.
• To train an attention-based, multi-task CNN for abnormality detection and location estimation in whole-body PET/CT scans.
• To improve performance on rare abnormalities and enhance mortality prediction using imaging data.

Main Methods:

• Leveraged natural language processing advancements to extract weak labels from free-text radiology reports.
• Developed a custom ontology for anatomical regions to structure extracted pathology information.
• Trained an attention-based, multi-task Convolutional Neural Network (CNN) architecture on the generated labels.

Main Results:

• The weak supervision framework successfully generated granular, structured abnormality labels from unstructured radiology reports.
• The multi-task CNN demonstrated strong performance in detecting and locating abnormalities, especially rare ones with limited data.
• The learned representation improved the accuracy of mortality prediction from PET/CT imaging data.

Conclusions:

• Weak supervision using NLP offers a viable solution to overcome data limitations in whole-body PET/CT analysis.
• The proposed framework enables more robust abnormality detection and has potential applications in prognostic modeling.
• This approach can significantly advance computational decision support in clinical workflows.