Limited-Angle CT Reconstruction Using Multi-Volume Latent Consistency Model
New multi-volume latent consistency model achieves 0.9677 SSIM with only 60-degree scan angles.
A research team from Kyoto University and Nagoya University has developed a breakthrough AI model for medical imaging that can reconstruct high-quality CT scans using significantly less data. Their Multi-Volume Latent Consistency Model addresses the critical challenge of limited-angle computed tomography (LACT), where traditional methods struggle with missing projection angles. The system combines 3D latent diffusion models with consistency models for fast inference, achieving remarkable results with just 60-degree scan angles while maintaining organ boundary precision and contrast that previous methods couldn't preserve.
The technical innovation lies in the multi-volume encoder that captures latent variables from different scales—both global regions and central regions—enabling stable performance across diverse clinical conditions. The model achieved MAE of 10.12 HU and SSIM of 0.9677 under 60-degree conditions, and maintained 0.9393 SSIM even at extreme 30-degree angles. Crucially, it demonstrated robustness with unknown projection angles not included in training, suggesting broad clinical applicability. This represents a significant advance over traditional diffusion models that struggled with 3D structure preservation and couldn't handle varying field-of-view conditions effectively.
- Achieves 0.9677 SSIM with only 60-degree scan angles (vs. full 180-degree traditional scans)
- Maintains 0.9393 SSIM at extreme 30-degree conditions with MAE of 16.69 HU
- Uses multi-volume encoder to handle different FOV conditions and preserves 3D organ structures
Why It Matters
Enables faster, lower-radiation CT scans while maintaining diagnostic quality, potentially revolutionizing medical imaging protocols.