[Paper Review] Parallel Diffusion Model-based Sparse-view Cone-beam Breast CT

Wenjun Xia, Tseng, Hsin Wu+7 more|arXiv (Cornell University)|Mar 22, 2023

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper proposes a parallel, dual-domain diffusion model framework for sparse-view cone-beam breast CT reconstruction, using separate Denoising Diffusion Probabilistic Models (DDPMs) trained on projection and image sub-volumes in parallel. The method achieves high-quality 3D breast CT images at half to one-third of standard radiation dose, significantly reducing dose while preserving diagnostic image quality.

ABSTRACT

Breast cancer is the most prevalent cancer among women worldwide, and early detection is crucial for reducing its mortality rate and improving quality of life. Dedicated breast computed tomography (CT) scanners offer better image quality than mammography and tomosynthesis in general but at higher radiation dose. To enable breast CT for cancer screening, the challenge is to minimize the radiation dose without compromising image quality, according to the ALARA principle (as low as reasonably achievable). Over the past years, deep learning has shown remarkable successes in various tasks, including low-dose CT especially few-view CT. Currently, the diffusion model presents the state of the art for CT reconstruction. To develop the first diffusion model-based breast CT reconstruction method, here we report innovations to address the large memory requirement for breast cone-beam CT reconstruction and high computational cost of the diffusion model. Specifically, in this study we transform the cutting-edge Denoising Diffusion Probabilistic Model (DDPM) into a parallel framework for sub-volume-based sparse-view breast CT image reconstruction in projection and image domains. This novel approach involves the concurrent training of two distinct DDPM models dedicated to processing projection and image data synergistically in the dual domains. Our experimental findings reveal that this method delivers competitive reconstruction performance at half to one-third of the standard radiation doses. This advancement demonstrates an exciting potential of diffusion-type models for volumetric breast reconstruction at high-resolution with much-reduced radiation dose and as such hopefully redefines breast cancer screening and diagnosis.

Motivation & Objective

• To address the high radiation dose in dedicated breast CT by enabling low-dose scanning while maintaining image quality.
• To overcome the large memory and computational cost of diffusion models in volumetric breast CT reconstruction.
• To develop a parallel, sub-volume-based framework that enables efficient training and inference on large 3D datasets.
• To integrate diffusion models in both projection and image domains for synergistic reconstruction performance.
• To validate the method on real sparse-view cone-beam breast CT data at reduced dose levels

Proposed method

• A dual-domain parallel pipeline is designed, training two separate DDPMs: one for projection data (DDPM-P) and one for image data (DDPM-I).
• The 3D projection and image volumes are divided into non-overlapping 16×16×16 sub-volumes for parallel processing.
• DDPM-P restores missing data in sparse-view projections using a U-Net-based architecture with 1,000 diffusion steps and a linear beta schedule.
• After sub-volume inpainting in the projection domain, the full dataset is reconstructed using the FDK algorithm.
• DDPM-I refines low-quality FDK-reconstructed image volumes by learning from paired low-quality and high-quality image sub-volumes.
• The entire pipeline leverages distributed computing with data broadcasting and assembly, enabling efficient inference on GPU clusters.

Experimental results

Research questions

• RQ1 Can a parallel, dual-domain diffusion model framework effectively reconstruct high-quality 3D breast CT images from sparse-view cone-beam projections?
• RQ2 How does the proposed parallel sub-volume processing strategy reduce memory and computational costs in diffusion-based CT reconstruction?
• RQ3 To what extent can the dual-domain DDPM approach maintain image quality at half to one-third of the standard radiation dose?
• RQ4 How does the synergy between projection-domain and image-domain DDPMs improve reconstruction fidelity compared to single-domain or conventional methods?
• RQ5 Can the method be efficiently deployed on a GPU-based cluster for clinical-scale reconstruction?

Key findings

• The proposed method achieves competitive image reconstruction quality at half to one-third of the standard radiation dose, as validated on real cone-beam breast CT data.
• The parallel sub-volume processing strategy significantly reduces memory requirements and enables efficient training and inference on large 3D datasets.
• The dual-domain DDPM framework outperforms conventional FDK reconstruction in terms of image quality, especially in reducing noise and artifacts in sparse-view data.
• The use of separate DDPMs for projection and image domains enables synergistic refinement, with DDPM-I effectively enhancing FDK-reconstructed images.
• The method was successfully trained and tested on the AiMOS GPU cluster using 1,576 V100 GPUs, demonstrating scalability and computational feasibility.
• The framework is deployable on cloud computing platforms, supporting future clinical integration for low-dose breast cancer screening.