Stationary CT Imaging of Intracranial Hemorrhage with Diffusion Posterior Sampling Reconstruction

TL;DR

This paper introduces a diffusion posterior sampling (DPS) method for 3D CT reconstruction that significantly improves image quality in sparse sampling scenarios, especially for intracranial hemorrhage detection.

Contribution

The study develops a novel DPS algorithm tailored for stationary CT systems with multi-x-ray sources, enhancing reconstruction robustness against noise and undersampling artifacts.

Findings

DPS reduces directional artifacts by ~130% compared to PWLS.

DPS improves lesion shape recovery by 30% (DICE coefficient).

Enhanced visualization of brain features in experimental sCT data.

Abstract

Diffusion Posterior Sampling (DPS) can be used in Computed Tomography (CT) reconstruction by leveraging diffusion-based generative models for unconditional image synthesis while matching the observations (data) of a CT scan. Of particular interest is its application in scenarios involving sparse or limited angular sampling, where conventional reconstruction algorithms are often insufficient. We developed a DPS algorithm for 3D reconstruction from a stationary CT (sCT) portable brain stroke imaging unit based on a multi-x-ray source array (MXA) of 31 x-ray tubes and a curved area detector. In this configuration, conventional reconstruction e.g., Penalized Weighted Least Squares (PWLS) with a Huber edge-preserving penalty, suffers from severe directional undersampling artifacts. The proposed DPS integrates a two-dimensional diffusion model, acting on image slices, coupled to sCT data consistency and volumetric regularization terms to enable 3D reconstruction robust to noise and incomplete sampling. To reduce the computational burden of DPS, stochastic contraction with PWLS initialization was used to decrease the number of diffusion steps. The validation studies involved simulations of anthropomorphic brain phantoms with synthetic bleeds and experimental data from an sCT bench. In simulations, DPS achieved ~130% reduction of directional artifacts compared to PWLS and 30% better recovery of lesion shape (DICE coefficient). Benchtop studies demonstrated enhanced visualization of brain features in a Kyoto Kagaku head phantom. The proposed DPS achieved improved visualization of intracranial hemorrhage and brain morphology compared to conventional model-based reconstruction for the highly undersampled sCT system.