Synthetic Breast Ultrasound Images: A Study to Overcome Medical Data Sharing Barriers

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Data privacy

Background:

• Medical big data holds significant potential for improving healthcare but is underutilized due to privacy concerns.
• These concerns limit cross-center data sharing and the creation of large, diverse datasets.
• This hinders the development of advanced AI models for medical diagnosis.

Purpose of the Study:

• To develop a privacy-preserving method for synthesizing medical data, specifically breast ultrasound (US) images.
• To overcome data sharing barriers and enable the construction of larger, more diverse medical datasets.
• To evaluate the utility of synthetic US images for augmenting training data in breast lesion classification.

Main Methods:

• Developed CoLDiT, a conditional latent diffusion model with a transformer backbone, for generating synthetic breast US images.
• Trained CoLDiT on a large dataset of 9,705 real US images from 5,243 patients across 202 hospitals.
• Validated the privacy of generated images using nearest-neighbor analysis and assessed realism through blinded reader studies.

Main Results:

• CoLDiT successfully generated realistic synthetic breast US images across various BI-RADS categories without duplicating private information.
• Blinded reader studies showed good performance in distinguishing real from synthetic images (AUC 0.53-0.77).
• Augmenting a BI-RADS classification training set with synthetic images yielded performance comparable to using only real images (P=0.81 for AUC).

Conclusions:

• Synthetic medical data generated by models like CoLDiT offers a viable, privacy-preserving solution.
• This technology can facilitate secure medical data sharing, overcoming current limitations.
• It enables greater utilization of medical big data to advance healthcare outcomes and AI development.