Machine Learning-Based Privacy Preserving via CT/MRI and Organ Metadata Prediction

Area of Science:

• Medical Imaging
• Machine Learning
• Data Privacy

Background:

• Medical imaging is crucial for diagnosis and treatment.
• Machine learning (ML) models require data for training.
• Patient privacy concerns necessitate robust data deidentification, often involving metadata removal.

Purpose of the Study:

• To develop an automated method for predicting essential metadata from fully anonymized medical images (CT and MRI).
• To enable efficient training of downstream ML models without compromising patient privacy.
• To integrate metadata prediction into privacy-preserving medical imaging workflows.

Main Methods:

• A novel framework combining machine learning and deterministic techniques was employed.
• The method predicts imaging modality (CT/MRI), anatomical region (heart, brain, liver), and MRI contrast type (T1/T2).
• The approach processes fully anonymized image data, inferring metadata directly from pixels.

Main Results:

• Achieved 100% accuracy in distinguishing CT/MRI modalities.
• Reached 99.2% accuracy in classifying anatomical regions (brain, heart, liver).
• Attained 99.8% accuracy in classifying MRI T1/T2 protocols.

Conclusions:

• The proposed framework successfully predicts technical metadata from anonymized medical images.
• This approach facilitates privacy-preserving medical imaging workflows without hindering downstream ML tasks.
• The method offers a practical solution for leveraging medical imaging data while upholding stringent privacy standards.