Limited-Angle CT Reconstruction Using Multi-Volume Latent Consistency Model

TL;DR

A multi-volume latent diffusion model that uses three-dimensional latent representations obtained from multiple effective fields of view as guidance for LACT reconstruction in clinical practical problems achieves fast and stable inference, and enables high-precision preservation of organ boundary information and internal structures under different FOV conditions.

Abstract

Limited-angle computed tomography (LACT) reconstruction is an inverse problem with severe ill-posedness arising from missing projection angles, and it is difficult to restore high-precision images without sufficient prior knowledge. In recent years, machine learning methods represented by diffusion models have demonstrated high image generation capabilities. However, accurate restoration of three-dimensional structures of organs and vessels and preservation of contrast remain challenges, and the impact of differences in diverse clinical imaging conditions such as field of view (FOV) and projection angle range on reconstruction accuracy has not been sufficiently investigated. In this study, we propose a multi-volume latent diffusion model that uses three-dimensional latent representations obtained from multiple effective fields of view as guidance for LACT reconstruction in clinical practical problems. The proposed method achieves fast and stable inference by introducing consistency models into latent space, and enables high-precision preservation of organ boundary information and internal structures under different FOV conditions through a Multi-volume encoder that acquires latent variables from different scales of the global region and central region. The evaluation experiments demonstrated that the proposed method achieved high-precision synthetic CT image generation compared to existing methods. Under the limited-angle condition of 60 degrees, MAE of 10.12 HU and SSIM of 0.9677 were achieved, and under the extreme limited-angle condition of 30 degrees, MAE of 16.69 HU and SSIM of 0.9393 were achieved. Furthermore, stable reconstruction performance was demonstrated even for unknown projection angle conditions not included during training, confirming the applicability to diverse imaging conditions in clinical practice.

Figures (9)

• Figure 1: Computed tomography (CT)/cone-beam CT (CBCT) reconstruction and limited-angle CT (LACT) image generation using generative artificial intelligence (AI). Synthetic CT (sCT) images obtained via generative AI-based LACT image completion remain challenging. Green arrows highlight incorrectly reconstructed voxels exhibiting low contrast and rough surface structures.
• Figure 2: Flow of limited-angle computed tomography (LACT) image simulation for self-supervised learning of image completion. $\theta_s$: angle at projection start, $\theta_r$: angular range from projection start to end and $\Delta \theta$: angular interval between each projection image.
• Figure 3: Self-supervised multi-volume conditional latent diffusion model for LACT image restoration. Multi-volume encoding extracts both global anatomical structures and high-frequency components through different convolution pathways from whole and local 3D regions. Restoration from noisy latents at two time points are learned during diffusion training. performed.
• Figure 4: Conceptual overview of the proposed Multi-Volume Vector Quantized Variational Auto Encoder architecture. The two latent representations ${z}_{G}$ and ${z}_{L}$ have the same spatial resolution but encode image features at different resolution levels in the image space.
• Figure 5: Impact of multi-view encoding on generated images. Impact of multi-view encoding on generated images. Bounding boxes and arrows highlight the improved visibility of organ and skeletal contours achieved with multi-view encoding.