Abstract
Cardiac Single-Photon Emission Computed Tomography (SPECT) is a critical tool for diagnosing coronary artery disease, yet it is often hampered by technical limitations: high image noise in low-dose protocols, artifacts from limited-view detector systems, and the requirement for CT-based attenuation correction. We propose DuDoCFNet, a dual-domain, multi-task deep learning framework designed to solve these challenges simultaneously. DuDoCFNet iteratively processes data in both the projection and image domains, performing denoising and limited-view reconstruction while synthesizing anatomical attenuation maps (μ-maps). Evaluated on 600 clinical studies, DuDoCFNet achieves state-of-the-art performance, offering a path toward high-quality cardiac SPECT imaging with reduced radiation exposure and simplified hardware requirements.
Introduction
Single-Photon Emission Computed Tomography (SPECT) imaging of the heart faces a "triple threat" of technical challenges that impact diagnostic quality and patient safety. First, to minimize radiation dose, clinicians often use Low-Dose (LD) protocols, which result in high levels of Poisson noise. Second, modern cost-effective scanners, such as the GE MyoSPECT ES, utilize Limited-View (LV) detector geometries to speed up acquisition, which introduces significant "missing data" artifacts. Third, accurate reconstruction requires an attenuation map (μ-map) to correct for photon absorption and scattering in tissue. Traditionally, this map is derived from a separate CT scan, which increases radiation dose and introduces potential misalignment issues.
While previous deep learning approaches have addressed these problems in isolation, DuDoCFNet is the first framework to treat denoising, limited-view reconstruction, and attenuation correction as a single, interrelated multi-task problem, leveraging the inherent synergy between raw data restoration and anatomical reconstruction.
Methodology
The core of DuDoCFNet is its Dual-Domain (DuDo) architecture, which facilitates information exchange between the Projection Domain (raw detector data) and the Image Domain (3D reconstruction).
Two-Stage Progressive Network (TSP-Net)
Operating in the projection domain, TSP-Net employs a coarse-to-fine strategy. The first stage focuses on restoring the general structure of the limited-view scan, while the second stage performs fine-grained denoising. This progressive approach ensures that the fundamental signal is recovered before noise suppression is applied.
Boundary-Aware Network (BDA-Net)
In the image domain, BDA-Net is responsible for predicting the μ-map. It incorporates a Spatial Boundary Enhancement (SBE) module. Since attenuation correction is highly sensitive to the boundaries between organs (e.g., heart-lung interfaces), the SBE module ensures these edges are reconstructed with high precision, which is critical for accurate final image correction.
Cross-Domain Integration
The network uses Multi-Layer Fusion (MLF) and Cross-Domain Feature Fusion (CDF) modules to bridge the two domains. This allows anatomical features from the image domain to guide the denoising of projections, and vice versa. Furthermore, the model incorporates Forward Projection (FP) and Back Projection (BP) operators, ensuring the learning process is constrained by the physical principles of SPECT imaging. The entire process is refined iteratively (typically 5 iterations) to progressively improve the quality of both the projections and the final reconstructed images.
Experimental Results
The framework was evaluated using a dataset of 600 clinical hybrid SPECT-CT studies. Low-dose (10% of original dose) and limited-view conditions (using 9 of 19 detectors) were simulated to test the model's robustness against ground-truth full-dose, full-view scans.
Quantitative Performance
DuDoCFNet demonstrated superior performance across multiple metrics compared to state-of-the-art models like U-Net and CDI-Net:
- Projection Restoration: The model achieved a Normalized Mean Square Error (NMSE) of 1.83% in the projection domain, outperforming the next best multi-task model (2.28%).
- Attenuation Correction: The generated μ-maps were highly accurate, particularly at tissue interfaces, with an NMSE of 11.43%.
- Clinical Utility: Reconstructed images showed a high correlation (0.9315) with ground-truth clinical polar maps, indicating that the synthesized μ-maps can effectively replace CT-derived maps for attenuation correction in many scenarios.
Efficiency and Synergy
A key finding was the synergy between tasks; denoising the raw data improved the 3D reconstruction, while the anatomical priors from the reconstruction enhanced the denoising process. Despite its iterative complexity, DuDoCFNet maintains an inference time of under one second, making it viable for real-time clinical use.
Conclusion and Future Work
DuDoCFNet represents a significant advancement in cardiac SPECT by providing a unified solution for noise reduction, artifact removal, and CT-less attenuation correction. By leveraging dual-domain processing and physics-informed constraints, it produces high-quality diagnostic images from sparse, low-dose data.
Future research will focus on improving the model's performance for anatomical structures outside the central cardiac region and integrating motion correction for respiratory and cardiac cycles. Additionally, further validation across different scanner types and tracers is necessary to ensure the generalizability of the framework in diverse clinical settings.
Summary updated