Date of Award
Spring 2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Biomedical Engineering (ENAS)
First Advisor
Liu, Chi
Abstract
Single-Photon Emission Computed Tomography (SPECT) is a non-invasive nuclear medicine imaging technique that visualizes the radiotracer distribution within the patient body through the detection of the emitted gamma-ray photons. Myocardial perfusion imaging (MPI) using SPECT is the most widely performed nuclear medicine exam that allows sensitive detection, localization, and risk stratification of ischemic heart diseases. Cardiac SPECT imaging presents a range of challenges and topics that are worth investigation, including attenuation correction (AC), limited-view (LV) reconstruction, cross-modality registration, low-dose denoising, etc. AC is essential for performing accurate quantitative or semi-quantitative analysis in cardiac SPECT. Currently, CT transmission scanning is typically used to generate attenuation maps (μ-maps) for the AC of cardiac SPECT. However, the additional CT scanning is limited by a series of issues including increased hardware expenses, accumulated radiation exposure, low market share of SPECT-CT hybrid scanners, etc. Additionally, the long scanning time is another problem hindering the quick application of SPECT MPI. Reducing the number of acquisition angles can reduce the total scanning time, but it also leads to degraded reconstruction accuracy due to reduced angular sampling. For cardiac SPECT-CT hybrid scanners, SPECT and CT scans are acquired sequentially, potentially inducing misregistration between the two images and further producing AC artifacts. In clinical practice, the SPECT-CT registration is manually performed by technologists, which is largely limited by the tedious manual procedures and inter-operator variabilities. Furthermore, reducing the dose of the injected radiotracer is essential for lowering the patient's radiation exposure in cardiac SPECT, but it will lead to increased image noise. Lastly, although various methods have been developed to solely focus on CT-free AC, LV reconstruction, or denoising, the solution for simultaneously addressing these tasks remains challenging and under-explored. Thus, in this thesis, we developed a series of deep learning techniques for the CT-free AC, LV reconstruction, SPECT-CT registration, and denoising, to promote and optimize the clinical application of cardiac SPECT imaging. Firstly, we developed a Dual Squeeze-and-Excitation Residual Dense Network (DuRDN) for generating the AC SPECT image from the non-attenuation-corrected (NAC) SPECT image in dedicated cardiac SPECT. DuRDN can enhance AC accuracy by incorporating the non-imaging patient information and the multi-energy-window imaging features. Then, we developed and compared the performance of the direct and indirect deep learning strategies for AC in both general and dedicated cardiac SPECT scanners. Moreover, we developed and validated the cross-vendor, cross-tracer, and cross-protocol deep transfer learning for the μ-map generation in SPECT MPI. Secondly, we proposed a Dual-Do}main Sinogram Synthesis (DuDoSS) method to predict synthetic full-view (FV) projections from sparsely-sampled LV projections. DuDoSS utilizes the images predicted in the image domain as the prior information to generate synthetic FV projections in the sinogram domain. The synthetic projections are then used for SPECT reconstruction. DuDoSS is the first dual-domain strategies used for LV reconstruction in nuclear medicine imaging. Thirdly, for the cross-modality SPECT-CT registration, we developed a Dual-Channel Squeeze-Fusion-Excitation (DuSFE) co-attention strategy. DuSFE effectively extracts and fuses the channel-wise and spatial information from SPECT and CT, largely improving the capability of encoding spatial features and thus enhancing the SPECT-CT registration performance. The performance of DuSFE was validated using both simulation data and real clinical application scenarios. Finally, we developed a Joint Dual-Domain Network (Joint-DuDo) for joint denoising and LV reconstruction of low-dose cardiac SPECT. Our proposed Adaptive Data Consistency (ADC) module improves the data fusion capability and thus enhances the prediction accuracy. Furthermore, we proposed a Dual-Domain Coarse-to-Fine Progressive Network (DuDoCFNet), a multi-task learning method for simultaneous LD denoising, LV reconstruction, and CT-free μ-map generation of cardiac SPECT. Paired dual-domain networks in DuDoCFNet are cascaded using a multi-layer fusion mechanism for cross-domain and cross-modality feature fusion, thus improving the prediction performance of each individual task.
Recommended Citation
Chen, Xiongchao, "Deep Learning Methods of Attenuation Correction, Reconstruction, Registration, and Denoising for Multi-Modal Cardiac SPECT" (2024). Yale Graduate School of Arts and Sciences Dissertations. 1457.
https://elischolar.library.yale.edu/gsas_dissertations/1457
