Sparseview SPECT Image Enhancement (SPASHT)

• Sparseview SPECT Image Enhancement (SPASHT) is a framework that combines inverse problem formulations with deep learning to reconstruct high-quality images from limited angular views.
• It employs model-based priors and task-aware loss functions to efficiently suppress streak artifacts and noise while preserving critical diagnostic features.
• Modern implementations using CNNs, diffusion models, and cross-domain networks demonstrate near full-view performance with significant reductions in scan time and patient dose.

Sparseview SPECT Image Enhancement (SPASHT) encompasses algorithmic and deep learning frameworks for reconstructing or enhancing single-photon emission computed tomography (SPECT) images acquired with far fewer angular views than standard protocols. The defining objective of SPASHT is to mitigate the streak artifacts, noise, and structural degradation that arise in this heavily underdetermined tomographic regime, enabling either shortened scan times or reduced patient dose without substantial loss of diagnostic information. SPASHT encompasses approaches ranging from principled inverse problem formulations with custom sparsity and blur priors to modern task-optimized deep neural architectures—including architectures augmented for generalization across diverse acquisition protocols, integration with anatomical priors, or explicit task-oriented loss functions such as observer models.

1. Sparse-View SPECT Imaging: Physical and Inverse Problem Formulation

In SPECT, the measurement process is typically modeled as a linear or linearized system: $y = A x + n$ where $y \in \mathbb{R}^M$ denotes the projection (sinogram) data for $M$ detector bins, $x \in \mathbb{R}^N$ is the target activity image (possibly 3D), $A \in \mathbb{R}^{M \times N}$ is the system matrix (describing geometry, collimator, attenuation, etc.), and $n$ is Poisson or Gaussian noise.

Under "sparse-view" sampling, the number of projections $V'$ is much less than the standard ($V$), with $M = V' \times P$ for $P$ detector bins per view, where $V'/V$ can be as small as 1/6. This underdetermined regime yields profound streaking and loss of high-frequency information when reconstructed with classical methods such as maximum-likelihood expectation maximization (MLEM) or ordered-subset expectation maximization (OSEM).

SPASHT approaches are formulated to address this ill-posedness via:

• Imposing strong prior knowledge on the solution class (e.g., sparsity in TV domain, piecewise-constant with blur, anatomical priors).
• Data-driven direct inversion through deep neural networks trained on large, realistic SPECT datasets.
• Explicit task-aware loss functions that preserve diagnostically relevant features or human observer performance.

2. Model-Based and Sparse Regularization Approaches

Early SPASHT algorithms exploit sparsity in specific transform or image domains while accounting for intrinsic blur due to physical effects. For example, the approach of Wolf et al. defines the object as a piecewise-constant "sharp" image $f$ convolved with a Gaussian blur, giving the final image $u = G M f$, where $G$ is a Gaussian operator and $M$ restricts the support to the FOV (Wolf et al., 2012):

$$\text{minimize}_f \left\{ D_{KL}(b, H G M f) + \lambda \| Df \|_1 \right\}$$

where $D_{KL}$ is the generalized Kullback-Leibler divergence and $D$ is the discrete gradient operator (total variation). Optimization uses primal-dual schemes, tuning parameters $\lambda$ and blur SD $r$ to control the tradeoff between artifact suppression (high $\lambda, r$) and spatial resolution (low $\lambda, r$).

Monte Carlo and "inverse-crime" simulations demonstrate that SPASHT achieves <3% loss in correlation coefficient (CC) compared to full-view, even at 9 angular views. SNR and artifact reduction are substantially improved over MLEM, but high-frequency textures not modeled by the prior can be oversmoothed.

3. Deep Learning Architectures for Sparse-View SPECT Enhancement

Modern SPASHT methods employ deep convolutional neural networks (CNNs) or diffusion models to perform direct denoising, artifact removal, and structural restoration in the sparse-view regime.

Canonical encoder–decoder approaches: For example, the method in (Chrysostomou et al., 2021) uses an hourglass convolutional network to invert the sparse projection operator, mapping noisy low-angle sinograms to reconstructed images. The architecture comprises a multi-stage encoder (Conv2D/LeakyReLU/BatchNorm/MaxPooling/Dropout) followed by a mirrored decoder (Conv2DTranspose/BatchNorm/LeakyReLU), with 15M–20M total parameters. Training uses a structural similarity (SSIM) loss. This approach achieves PSNR ≈25 dB and SSIM = 0.92 at 24 views, outperforming MLEM by a significant margin across both simulated and hardware phantoms.

3D U-Net–style networks (for task-oriented performance): SPASHT frameworks tailored for myocardial perfusion imaging (MPI) SPECT (e.g., (Yang et al., 9 Nov 2025, Yang et al., 22 Apr 2025)) use 3D encoder–decoder networks (Conv3D/BatchNorm/ReLU/Dropout, skip connections) operating directly on short-axis cardiac volumes. The input is an OSEM reconstruction from 5–15 sparse projections; the output is a volume that approximates the full-view reconstruction.

Notably, SPASHT integrates anthropomorphic observer loss terms—specifically, the channelized Hotelling observer (CHO) channel matrix $U$ operating on 2D slices—into the training objective: $L_{\rm total} = L_{\rm rec} + \lambda L_{\rm obs}$ where $L_{\rm rec}$ is voxel-wise MSE and $L_{\rm obs}$ penalizes differences in channel responses between predicted and full-view images, explicitly targeting diagnostic task performance.

Task-agnostic deep learning methods are outperformed by these observer-model-driven losses, with observed area under the ROC curve (AUC) improvements from 0.75 (sparse-view) to 0.88 (SPASHT) at 5 views, and similar gains at more moderate sub-sampling rates.

4. Generalizable and Task-Specific Deep Architectures

Diffusion models and 2.5D conditioning: The "DiffSPECT-3D" framework (Xie et al., 21 Dec 2024) extends SPASHT via stochastic diffusion models robust across multiple acquisition settings without retraining. Here, each diffusion timestep denoises a 2D slice conditioned on the entire 3D cardiac CT, using positional encodings for temporal and spatial information. Data consistency is enforced:

• In the projection domain, via periodic MLEM update steps,
• In the image domain, via gradient steps toward under-sampled MLEM reconstructions,
• With a total variation (TV) penalty along the z-axis.

The model is trained by denoising-score matching: $$\mathcal L_{\rm denoise} = \mathbb E_{x_0,\epsilon\sim\mathcal N(0,I),\,t}\left[\left\|\epsilon-\epsilon_\theta(x_t,t)\right\|^2\right]$$ and inference involves alternating diffusion-sampling, data-consistency, and TV-proximal updates (25 DDIM steps typical, ≈2 s/volume). DiffSPECT-3D achieves high quantitative fidelity (SSIM = 0.946 at 9/19 views) and preserves diagnostic accuracy in clinical readouts, even at aggressive sparse-view or low-dose settings.

Cross-domain iterative networks (CDI-Net) (Chen et al., 2023): This framework jointly estimates emission and attenuation maps via a sequence of coupled U-Nets, passing representations between the projection and image domains at each iteration. An Adaptive Weight Recalibrator fuses heterogeneous feature maps, and loss functions combine projection, attenuation, and image-domain supervision. CDI-Net matches or exceeds prior state-of-the-art approaches in noise suppression, structural recovery, attenuation correction, and diagnostic task performance.

5. Quantitative Performance, Clinical Evaluation, and Task-Oriented Metrics

Quantitative performance of SPASHT algorithms is assessed using:

• Standard metrics: PSNR, SSIM, RMSE, Pearson correlation coefficient.
• Task-specific metrics: area under the receiver operating characteristic curve (AUC) for perfusion-defect detection using channelized Hotelling observers or human readers.

Sparse-View Level	Sparse (AUC)	Task-Agnostic DL (AUC)	SPASHT (AUC)	Full-View (AUC)
1/6 (5 views)	0.75 ± 0.03	0.78 ± 0.02	0.87 ± 0.02*	0.98 ± 0.01
1/3 (10 views)	0.82 ± 0.02	0.84 ± 0.02	0.92 ± 0.01*	0.99 ± 0.01
1/2 (15 views)	0.88 ± 0.02	0.89 ± 0.01	0.96 ± 0.01*	0.99 ± 0.01

(*p < 0.05 vs. sparse-view; (Yang et al., 9 Nov 2025))

SPASHT frameworks have consistently demonstrated large gains in both model-observer and human-reader ROC AUCs relative to conventional sparse-view methods, narrowing the diagnostic performance gap to full-view protocols by over 75%.

Clinical studies (Yang et al., 9 Nov 2025, Xie et al., 21 Dec 2024) indicate that SPASHT enhancement supports scan-time reductions by factors of 2–6 with little loss in diagnostic accuracy, reduces motion artifacts, and enables protocol flexibility.

6. Practical Considerations, Limitations, and Extensions

Resource and implementation requirements:

• Model-based inverse algorithms: substantial memory, Ch-Pock primal-dual solvers, per-protocol parameter sweeps (TV weight $\lambda$, blur SD $r$), extension to 3D involving further memory overhead (Wolf et al., 2012).
• Deep learning: substantial computational resources for training (e.g., 500,000 sample datasets, 10–24 hr single-GPU training), inference times as low as 10 ms per 128×128 image or ≈2 s per 3D volume on suitable hardware (Chrysostomou et al., 2021, Xie et al., 21 Dec 2024).

Limitations:

• Many frameworks are trained/tested on synthetic or simulated defects; broader validation on clinical, invasive reference standards is required (Yang et al., 9 Nov 2025).
• Current SPASHT models may “hallucinate” features outside their training manifold or fail to generalize if acquisition geometry, patient population, or scanner varies significantly.
• Most task losses focus on only a subset of anatomical/perfusion defect types; additional regions and multi-task objectives may be necessary for clinical adoption.

Extensibility:

• Incorporation of anatomical priors from CT or MRI is now routine in diffusion or cross-domain networks, improving anatomical plausibility and attenuation correction (Xie et al., 21 Dec 2024, Chen et al., 2023).
• Hybrid loss functions, e.g., SSIM plus total variation or observer loss, further improve recovery of both global structure and task-relevant features.
• Architectures can be readily adapted to related modalities (e.g., PET, low-count/few-view CT) and extended for domain adaptation or multi-center generalization.

7. Future Directions

Recent SPASHT studies propose several directions for further research:

• Multi-center, multi-vendor validation for generalizability, especially in clinical defect detection with unknown target locations (Yang et al., 9 Nov 2025).
• Joint dose- and view-reduction protocols leveraging SPASHT frameworks.
• Physics-driven and frequency-domain analyses to dissect and optimize observer-channel losses, interpretability, and diagnostic robustness (Yang et al., 9 Nov 2025, Xie et al., 21 Dec 2024).
• Unsupervised or semi-supervised domain adaptation to mitigate dependency on co-registered anatomical maps (Chen et al., 2023).
• Development of scalable 3D diffusion or transformer-based models that can operate on full cardiac volumes end-to-end without expensive slicing or severe memory constraints.

A plausible implication is that integration of anatomical information, explicit task loss, and generalization mechanisms will become standard components of future SPASHT pipelines, with clinical trials needed to fully establish their efficacy and safety in practice.