Speaker
Description
Background: Alpha-particle radiopharmaceutical therapy (α-RPT) is a promising treatment modality for cancers such as metastatic castration-resistant prostate cancer. Compared to β-emitters (e.g. ¹⁷⁷Lu), α-emitters such as ²²⁵Ac deliver high linear energy transfer over a short range, enabling effective tumour cell killing while sparing surrounding tissue. However, clinical translation is limited by the lack of reliable quantitative imaging and dosimetry. SPECT imaging in α-RPT is particularly challenging due to low administered activities, resulting in highly ill-conditioned reconstruction problems with very limited photon counts. Conventional methods such as MLEM are highly sensitive to noise and regularized approaches are often insufficient while many AI-based methods often fail under such extreme conditions.
In this work, we propose a fully three-dimensional diffusion with posterior sampling (3D-DPS) framework for low-count SPECT reconstruction aiming to improve image quality while keeping data fidelity.
Methods: Our method leverages a generative diffusion model trained on 3D PSMA PET data to provide a population-based prior for radiotracer distribution. Unlike conventional DPS approaches based on l2 norm minimization, we formulate the data-consistency term using a Poisson log-likelihood consistent with standard iterative reconstruction methods for emission tomography such as MLEM/OSEM.
The diffusion model was trained using a variance-exploding stochastic differential equation with 1000 steps on 120 PSMA PET volumes, generating 3D activity distributions (128×128×128). During reconstruction, the learned prior is combined with a physics-based forward SPECT model through a data-consistency term.
Validation was performed using realistic SPECT simulations incorporating attenuation, collimator blur, and detector resolution. Five patient cases were simulated using 30 projection angles and ultra-low-count Poisson noise (~10⁵ detected photons). The proposed method was compared to MLEM and MLEM with total variation (TV) regularization.
Results and discussion: The proposed 3D-DPS significantly improves reconstruction quality compared to baseline methods. Quantitatively, we observed lower reconstruction error (NMAE = 0.36 ± 0.05) compared to MLEM (0.53 ± 0.02) and MLEM+TV (0.41 ± 0.05), corresponding to relative error reductions of ~32% and ~12%, respectively. Structural similarity is also improved (SSIM = 0.89 ± 0.05) compared to MLEM (0.77 ± 0.13) and MLEM+TV (0.85 ± 0.08), indicating better preservation of image features. Visually, the method suppresses noise while maintaining contrast and structural detail, whereas MLEM is dominated by noise and TV regularization leads to over smoothing and insufficient regularization for ultra low count scenarios.
Note that these improvements come with a preserved consistency with the measured data, with projection-domain log-likelihood values within 2% of standard MLEM reconstruction. This supports the assumption that the gains are not achieved at the expense of data fidelity, but rather through improved regularization driven by the learned prior.
Conclusions: In conclusion, the proposed framework provides a robust and physically consistent approach for reconstructing ultra-low-count SPECT data in α-RPT. This method has the potential to improve quantitative imaging and enable more reliable dosimetry in challenging low-count scenarios. Future work will focus on validation using Monte Carlo simulations and clinical datasets.
| Track | PSMR |
|---|---|
| Presentation type | Oral |