Lead-212 alpha-emitting radiopharmaceutical therapy has gained interest for tumor treatments. 212Pb emits one alpha particle in the decay chain for therapy and provides two imageable photon peaks – 77.4 keV and 238.6 keV – that allow image-based dose assessment using planar or single-photon emission computed tomography (SPECT). However, direct imaging of 212Pb is challenging due to down-scattered photons emitted by its daughter isotopes and a limited administered activity. These difficulties motivate the current development of elementally matched imaging surrogate 203Pb. While ensuring the same biodistribution as 212Pb, 203Pb provides 73.1 keV and 279.2 keV photon peaks with fewer higher-energy contaminants in the spectrum and allows a higher administered activity. In this study, we present a simulation workflow to evaluate planar imaging of a mock 212Pb/203Pb theranostic pair in a human phantom. The framework incorporates an analytical pharmacokinetic-decay model, organ-level dosimetry, and GPU-based Monte Carlo imaging for both 212Pb decay chain and 203Pb using different collimators and energy windows. Along with 203Pb, time-dependent biodistribution and photon effects of 212Pb, 212Bi, and 208Tl are simulated separately to identify the physical sources of image degradation in 212Pb imaging. The advantages of using 203Pb are quantified in terms of a better scatter fraction and lesion SNR within a specified imaging duration. Our simulation provides a platform that benefits in evaluating feasibility and performance of imaging alpha-radiopharmaceutical therapy with surrogates for clinical application.
Model-Based Deep Learning Iterative Reconstruction methods, particularly deep unrolled networks, are actively investigated for 3D Single Photon Emission Computed Tomography (SPECT) reconstruction. However, translating these methods to fully 3D clinical settings remains bottlenecked by the prohibitive memory required to store computational graphs. To address this, we propose a new training method for 3D SPECT called Jacobian-Free Unrolling (JFU). We unroll the Maximum A Posteriori Expectation Maximization (MAPEM) algorithm, parameterizing the gradient of the regularization term with a convolutional neural network trained on a synthetic dataset in a supervised setting. JFU addresses these memory requirements of 3D unrolling by backpropagating solely through the final inner iteration, which significantly reduces VRAM usage. We evaluated JFU on simulated and experimental 177Lu data, comparing it against several state-of-the-art baselines: OSEM with Resolution Modelling, MAPEM using Relative Difference Penalty prior, PVCNet-Hybrid and classical unrolling. JFU-MAPEM demonstrated superior training stability and improved image quality, outperforming these baselines across global and local metrics. On an experimental IEC phantom, JFU-MAPEM achieved higher Recovery Coefficients, approaching 1.0 for the two largest spheres compared to 0.7 to 0.9 for OSEM and MAPEM. Furthermore, visual assessment of clinical patient image confirmed that JFU-MAPEM yields smoother and realistic reconstructions. Finally, reconstruction with JFU-MAPEM was faster than standard MAPEM. Although the model relies on camera-specific synthetic training data, it provides a computationally efficient and promising framework for quantitative SPECT tasks. This approach is particularly valuable for accurate voxel-based dosimetry in Radiopharmaceutical Therapies and can be adapted to other Poisson-distributed inverse problemst.
>Aim: This study was designed to examine the effects of oxygen plasma surface treatment as well as nanodiamond surface deposition on 3D-printed PCL scaffolds. Methodology: 3D-printed PCL scaffold specimens with particular porosity were fabricated using the fused deposition molding (FDM) technique. The specimens were equally divided into three main groups according to the surface treatment: Control group (untreated specimens), Plasma-treated group (specimens treated with oxygen plasma), and Nanodiamonds group (specimens treated with nanodiamond deposition). The three groups were evaluated through morphological characterization by field emission scanning electron microscopy (FESEM), microstructure characterization by Raman spectroscopy, chemical characterization by Fourier transform infrared (FTIR) spectroscopy, hydrophilicity degree by contact angle, and in vitro degradation measurements (n=6). Also, qualitative and quantitative in vitro cell adhesion and osteogenic differentiation of bone marrow mesenchymal stem cells (BMSCs) (n=6) were evaluated. The statistical significance level was set at p ≤0.05. Results: The morphological characterization revealed an etched surface on oxygen plasma-treated specimens and evenly distributed nanoscale characteristics across the surfaces of the nanodiamond-treated specimens. The Raman spectra confirmed the successful surface deposition of nanodiamonds on the scaffolds, while the FTIR results confirmed the presence of oxygen-containing functional groups on the treated specimens. The nanodiamond-treated specimens showed the statistically significant lowest contact angle (19.85°, ±1.37). On the other hand, oxygen plasma-treated specimens revealed the greatest degradation rate across all incubation periods in phosphate-buffered saline (PBS). Moreover, the nanodiamond-treated scaffolds exhibited better quantitative and qualitative cell adhesion, in addition to better osteogenic differentiation compared to control and plasma-treated groups. Conclusion: Nanodiamond deposition on 3D-printed PCL scaffolds significantly increased their hydrophilicity, improved in vitro cell adhesion as well as osteogenic differentiation, compared to both untreated and oxygen plasma-treated 3D-printed PCL scaffolds.
Data-driven approaches to cardiac gating offer a means to reduce cardiac contractile motion blurring without the need for an external device. In this work, we extend existing methods and present a data-driven cardiac contractile gating approach for 82-Rb myocardial perfusion imaging (MPI). An anatomy based initial signal estimate is used to group voxels exhibiting oscillatory motion in short frame histo-images which have had respiratory motion removed by a previously validated method. A signal enhancement step further groups voxels by their phase to reduce destructive phase effects and reduce noise. Fourier analysis is then performed to both filter the signal and select the best signal at any point in the scan which accounts for effects of tracer kinetics in MPI. 15 patient rest and stress scans were used to test the methodology and the extracted signals were compared to ground truth ECG signals used clinically for cardiac gating. Direct comparison of the data-driven signal with the ECG R-wave peak time points revealed a mean cycle length error of 4.0%, 3.3%, 3.7% and 3.1% and a mean error in the standard deviation of trigger points of 7.2%, 4.7%, 6.7%, and 4.1% for the rest bolus, rest perfused, stress bolus and stress perfused portion of the scans respectively. Cardiac functional analysis showed a strong match between the ECG and data-driven approaches where a mean absolute error of 1.9 ml, 1.8 ml and 1.8% was found for the calculation of end diastolic volume, end systolic volume and left ventricle ejection fraction respectively. The proposed methodology advances previously presented methods for data-driven cardiac gating and expands the approach to dynamic myocardial perfusion imaging with 82-Rb. High performance on real world data suggests this approach may allow for device-less cardiac gating in real world scenarios.
Multi-phase Coronary Computed Tomography Angiography (CCTA) enhances clinical assessment by increasing diagnostic confidence through expanded visualization of cardiac phases and uncovering coronary motion dynamics. However, coronary motion artifacts in CCTA images increase the difficulty of coronary lesion identification and even lead to misdiagnosis. For artifact reduction, traditional physics-model-based motion correction methods show limited performance, while deep learning methods effectively mitigate artifacts but easily introduce unrealistic coronary structures, compromising imaging fidelity. To overcome these limitations, we introduce sectorized Limited Angle Reconstruction based Temporal Attention Dense Network (SLAR-TADenseNet) for fidelity-constrained coronary motion artifact reduction. Multi-phase SLAR images present inter-phase structural similarity, motion continuity, and explicit phase complementarity. Using SLAR images as model inputs provides physically grounded priors via their synergistic properties. The TADenseNet is designed to exploit these physically grounded priors from SLAR inputs to improve CCTA imaging fidelity. SLAR-TADenseNet integrates a SLAR-adapted Dense Connection Convolution module for deep feature reuse and multi-phase feature fusion, and a Temporal Self-Attention module for precise capture of coronary motion structures. The two modules form a synergistic mechanism to respectively handle static structural feature fusion and dynamic coronary motion capture. Simulation and clinical results demonstrate that SLAR-TADenseNet achieves superior artifact reduction and coronary fidelity while producing clear coronary images over multiple cardiac phases.
