Design and Analysis of Low-Cost X-Ray Imaging System Incorporating Add-On Mouse Rotation System in Pre-Clinical Research

Abstract

Computed tomography (CT) has become an essential tool in diagnostic and sports medicine. It is able to generate a three-dimensional rendering that can be used to interrogate a given region or desired structure from any orientation. However, in pre-clinical research, its deployment remains limited due to relatively high upfront costs. Besides, existing integrated imaging systems that provide X-ray modalities also dwarf its popularity in small laboratories. This lag of pre-clinical use of CT has resulted in small academic labs having to run pre-clinical CT scans on clinical veterinary scanners.

This dissertation seeks to generate CT images using an existing small-animal X-ray imager with a specialized add-on specimen rotation system, the \textit{MiSpinner}. This setup conforms to the cone-beam computed tomography (CBCT) geometry, which requires high spatial accuracy. The key problem in retrofit studies is that the relative position of portable add-on devices varies by case. Therefore, in this research, we pursue three aims based on projection-based image processing methods and neural network methods: 1. System geometry calibration, 2. Low-dose projection denoising, and 3. sparse sampling enrichment.

System geometry calibration is achieved by our proposed structure tensor-based two-step online (ST-TSO) calibration method. It combines with a comprehensive edge detection technique to detect spatial misalignment information on projections. Specifically, the first processing step employs a modified projection matrix-based calibration algorithm to estimate the relevant geometric parameters, and the second processing step fine-tunes the parameters with an iterative strategy based on the symmetry property of the sum of projections. We showed that by using this calibration method we can successfully reconstruct the image, and its accuracy outperforms other online calibration methods.

Low-dose projection image denoising is achieved with a model-based iterative reconstruction (MBIR) method. We consider the noise generation statistical model under a maximum a posterior probability (MAP) framework with a structure tensor based total directional variation (ST-TDV) image prior regularization term. We demonstrated that the proposed method can effectively remove the noise in low-dose cases using a digital phantom study.

Sparse sampling introduces extra streak artifacts into reconstructed images that degrade the image quality. We propose an adversarial learning-based spatial transformer network for projection image synthesis, which aims to improve the quality of reconstructed images by reasonably increasing the number of projection images based on the existing projection data. Simulation and experiments have shown that this data-driven model can give rise to competitive results compared with conventional algorithms.

Additionally, due to the lack of electron density reference standards, we have tried an image style transfer network, cycle generative adversarial network (cycle-GAN), to improve the reconstructed image contrast. The deployed network that only requires two groups of image datasets instead of image pairs has shown promise in retrofit studies, and our testing has confirmed its promising competitive performance.