Development of 3D patient-based super-resolution digital breast phantoms using machine learning.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Artificial Intelligence

Background:

• Digital phantoms are crucial for evaluating X-ray imaging systems.
• Existing phantoms often lack the anatomical detail and resolution needed for accurate human anatomy modeling.
• Patient-specific phantoms require high spatial resolution to capture anatomical nuances.

Purpose of the Study:

• To propose a pipeline for enhancing the spatial resolution of patient-based digital breast phantoms.
• To generate super-resolution phantoms for computer simulations in breast imaging.
• To improve the realism and detail of digital phantoms for research.

Main Methods:

• Developed a supervised learning pipeline to increase the spatial resolution of tomographic breast images.
• Algorithms predict and recover glandular details lost due to limited imaging resolution.
• Trained models on high-resolution synchrotron images and applied them to clinical breast CT data.

Main Results:

• Generated super-resolution digital breast phantoms with improved detail (68 μm voxel size).
• Achieved high accuracy in recovering glandular details (synchrotron: 0.95 ± 0.04; clinical: 0.15% ± 0.12% error).
• Radiologist evaluation confirmed the realism of the generated phantoms, indistinguishable from original images.

Conclusions:

• The proposed method successfully generates super-resolution digital breast phantoms from patient data.
• These phantoms are suitable for computer simulations to optimize new breast imaging technologies.
• This advancement enhances the utility of digital phantoms in medical imaging research.