SADIR: Shape-Aware Diffusion Models for 3D Image Reconstruction.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• 3D image reconstruction from limited 2D views is challenging.
• Current deep learning methods struggle with preserving object topology and shape structure.
• Artifacts like holes and discontinuities often occur in reconstructions.

Purpose of the Study:

• To propose a shape-aware network for improved 3D image reconstruction.
• To address limitations in topology preservation and artifact reduction.
• To leverage shape priors for more accurate 3D image generation.

Main Methods:

• Developed SADIR, a shape-aware network utilizing diffusion models.
• Integrated shape priors learned from training data to guide reconstruction.
• Employed a joint learning network to learn a mean shape under deformation models.

Main Results:

• SADIR demonstrated superior performance compared to baseline methods.
• Achieved lower reconstruction error in experiments.
• Showcased enhanced preservation of object shape structure.

Conclusions:

• The proposed SADIR model effectively reconstructs 3D images with improved shape and topology preservation.
• Leveraging shape priors within diffusion models offers a promising direction for 3D reconstruction.
• SADIR shows potential for applications in medical imaging, such as brain and cardiac MRIs.