A comparative study of diffusion-based reconstruction frameworks for photoacoustic tomography Authors: Shibili Said, Imad Barhumi Published: 2026 ABSTRACT AbstractPhotoacoustic tomography (PAT) reconstruction is a highly ill-posed inverse problem, particularly under limited-view acquisition, that leads to severe artifacts and loss of structural information. Diffusion-based generative models have recently been explored as a means of incorporating learned priors into image reconstruction. In this study, we systematically compare three representative diffusion-based frameworks: denoising diffusion probabilistic models (DDPM), denoising diffusion implicit models (DDIM), and score-based models for PAT reconstruction under full-view and limited-view conditions using both synthetic and anatomical data. Further, the role of physics-informed measurement-consistency refinement applied during sampling and its impact on reconstruction fidelity across varying degrees of data is analyzed. The results indicate that while DDPM provides accurate reconstructions under full-view acquisition, its performance degrades under limited-view conditions without an explicit refinement step. Similarly, DDIM achieves comparable reconstruction quality while requiring fewer sampling steps. In contrast, score-based models demonstrate consistently better performance across acquisition settings, yielding improved structural fidelity and perceptual quality, at the expense of a higher number of sampling timesteps. These findings provide practical insight into different diffusion formulations and offer guidance for deploying diffusion-based reconstruction methods in challenging PAT imaging scenarios.Keywords: Diffusion model, Photoacoustic tomography, Inverse problemSubject terms: Medical research, Optics and photonics Introduction Photoacoustic tomography (PAT) is an imaging technique that integrates optical excitation with ultrasonic detection. The short laser pulses are absorbed by tissue, leading to localized expansion and the emission of acoustic waves. These waves are detected and used to reconstruct the images enabling high-resolution visualization of tissue structure alongside functional and molecular information. Beyond vascular imaging, PAT provides insights into bone structure and composition and has applications in fracture assessment, osteoporosis evaluation, tumor detection, and imaging of joint and inflammatory diseases1. The other applications include functional brain imaging, melanoma detection, and the assessment of cardiovascular and inflammatory diseases2. The PAT technique is non-ionizing, cost-effective, and suitable for handheld use, making it ideal for bedside, intraoperative, and repeated imaging applications3. Photoacoustic image reconstruction is inherently an inverse problem, aiming to recover high-quality images from indirect and often noisy or incomplete measurements. Hence, robust reconstruction is essential to reconstruct anatomically meaningful photoacoustic images. In PAT, several factors limit reconstruction fidelity that can affect diagnostic reliability. The reconstruction task constitutes an ill-posed inverse problem, in which small measurement errors or modelling inaccuracies can result in substantial deviations in the recovered image4. Hence in practical PAT systems, these challenges aggravate reconstruction artifacts, making robust recovery from sparse or noisy measurements particularly challenging. Traditional iterative reconstruction methods are computationally expensive and sensitive to incomplete data, requiring many iterations and their reliance on simplified models and hand-crafted regularization often leads to unrealistic artifacts in noisy conditions5–7. Ongoing research demonstrates that diffusion models as powerful tools for solving inverse problems, especially in image reconstruction. Chung et al.8 proposed direct posterior sampling to solve general inverse problems, while Peng et al.9 demonstrated that learning an optimal posterior covariance in a plug-and-play framework substantially improves zero-shot diffusion-based solvers. In medical imaging, manifold-constrained gradient correction was introduced to keep sampling trajectories close to the data manifold, which improved reconstruction quality in computed tomography10. Bi-level Guided Diffusion Models further enhanced data fidelity through combined proximal optimization and posterior estimation11. Additional strategies, including pseudoinverse guidance and embedding forward models into each denoising step, highlight diverse approaches to measurement conditioning12,13. More recently, hard data consistency integrated into latent diffusion models has also enabled fast and high-fidelity reconstructions from corrupted CT measurements14. Diffusion models learn the underlying distribution of medical images and iteratively denoise to recover a high-quality image. By incorporating rich learned priors, diffusion-based reconstructions produce images with much higher fidelity and fewer distortions even from limited or corrupted measurements15,16. Emerging studies indicate that diffusion models have progressed significantly in recent years, evolving from their basic formulations into a wide range of variants tailored for diverse applications. In the context of PAT, reconstruction using diffusion models has not been explored enough, and existing studies focus on specific reconstruction strategy. A significant study demonstrated that score-based diffusion models excel at regularizing ill-posed inverse problems by leveraging data-driven priors, as in17 and further demonstrated on PAT in18. This has paved the way for the possible integration of diffusion models in solving inverse problems in PAT. Recent papers confirm that diffusion models can address key challenges in complex, ill-posed imaging systems, including noise and sparsity19–21. However, the relative behaviour of different diffusion formulations, such as probabilistic, deterministic, and score-based models, under identical photoacoustic reconstruction settings remains insufficiently explored. This study presents a comparative investigation of photoacoustic tomography reconstruction using three representative diffusion-based generative frameworks: denoising diffusion probabilistic models (DDPM), denoising diffusion implicit models (DDIM), and score-based generative models. The work conducted a unified, physics-consistent evaluation of probabilistic, deterministic, and continuous-time diffusion formulations for PAT. It provided a systematic assessment of their reconstruction behavior in full-view and limited-view acquisition using synthetic and anatomical datasets. In addition, the impact of an explicit refinement step in the sampling algorithm is examined to further clarify its role in improving the quality of reconstruction. Through this systematic evaluation, the study aims to provide insight into different diffusion-based formulations for PAT inverse problems. Method This study employed two fundamental diffusion frameworks, namely Denoising Diffusion Probabilistic Models (DDPM) and Denoising Diffusion Implicit Models (DDIM), and extended them to address the inverse reconstruction problem in photoacoustic tomography (PAT). Both DDPM and DDIM have demonstrated strong performance in generative sampling22,23. This study adopted the framework for PAT by integrating diffusion-based denoising with explicit enforcement of measurement consistency during the sampling process. Rather than using the diffusion model solely as a prior, consistency with the measured acoustic pressure data is enforced at each iteration, conceptually aligning the approach with score-based reconstruction strategies18. Reconstruction performance was systematically evaluated under both full-view and limited-view settings to analyze the impact of data incompleteness on image quality. Furthermore, experiments were also conducted with and without an additional refinement step following the measurement consistency update. Refinement stage is a gradient-based correction step that allows further reduction of residual discrepancies between the reconstructed images and the measured data. Comparative analysis of these configurations provided insight into the contribution of refinement to reconstruction accuracy and robustness. Overall, coupling stochastic diffusion with physics-based constraints, as illustrated in Algorithm 1, offers a practical approach for examining reconstruction fidelity in challenging PAT inverse problems. Reconstruction framework for DDPM and DDIM. Denoising diffusion probabilistic model (DDPM) DDPM, introduced by Ho et al.22, defines a class of generative models that learn data distributions through a sequence of gradual noise perturbations and learned denoising steps. In the forward process, Gaussian noise is progressively added to the data over a fixed number of time steps according to a predefined variance schedule, transforming the data into an approximately Gaussian distribution. The reverse process is modelled as a Markov chain, in which a neural network, typically a U-Net, is trained to predict and remove the added noise at each step, thereby reconstructing the data sample in a stepwise manner. During inference, samples are generated by starting from pure noise and iteratively applying the learned denoising steps. The stochastic reverse diffusion step in DDPM, where Gaussian noise is injected at each timestep to sample from the learned posterior distribution, as in (1). where, and denote the variables at diffusion steps t and t-1, respectively, denotes the cumulative product of the noise schedule, approximates the Gaussian noise added during the forward diffusion process, controls the amount of stochasticity injected at each reverse diffusion step, while denotes a standard Gaussian noise. Denoising diffusion implicit model (DDIM) Denoising Diffusion Implicit Models (DDIM)23 reformulate diffusion-based generative modeling by introducing a deterministic mapping between noise and data while preserving the same marginal distributions as Denoising Diffusion Probabilistic Models (DDPMs). Similar to DDPM, DDIM employs a forward diffusion process that progressively adds Gaussian noise according to a predefined variance schedule. The key distinction lies in the reverse process: where DDIM constructs a deterministic, non-Markovian reverse trajectory parameterized by the model’s noise prediction and a controllable stochasticity term. When this parameter is set to zero, the reverse process becomes fully deterministic, producing a unique reconstruction for a given noise initialization. This formulation enables sampling along shortened diffusion trajectories, allowing significantly faster inference without modifying the trained model. The DDIM reverse update is represented as (2). The parameter controls the amount of randomness introduced at each step of the reverse diffusion process. When is non-zero, Gaussian noise is added during sampling, causing the reverse trajectory to be stochastic. Setting removes randomness and results in a fully deterministic update as in DDIM. This deterministic formulation enables DDIM to achieve high-quality reconstructions using significantly fewer sampling steps, distinguishing it from standard DDPM sampling, which injects noise at each reverse step. Score based model Score-based models constitute a class of generative methods that learn the score function, defined as the gradient of the log probability density of the data, and use this information to guide the gradual removal of noise from corrupted samples. In the framework introduced by Song et al.17, score-based generative modelling is formulated as continuous time using stochastic differential equations (SDEs). The forward process progressively perturbs the data through a noise-driven diffusion process, while the reverse process leverages the learned score function to drive the samples back toward the data distribution. The model performs sampling by integrating the reverse-time stochastic differential equation as (3), where sampling is guided by the learned score function . where f(x, t) corresponds to the drift term describing how the state evolves in the absence of noise and g(t) is the diffusion coefficient that controls the strength of the Brownian noise , which injects randomness into the trajectory. These terms define a time-dependent distribution , where larger values of the diffusion time t correspond to progressively noisier versions of the state . Equation (3) is proceeded by using numerical solvers like predictor–corrector schemes that alternate between denoising updates and controlled stochastic refinement. In this work, we adopt a score-based reconstruction approach inspired by18 for PAT. The same model is further evaluated on multiple anatomical images to assess its generalization capability across different acquisition settings. Dataset The synthetic dataset18, consisting of branch-like structures that imitate vascular morphologies as seen in photoacoustic tomography (PAT), was used across all experiments to train the models. It includes 10,000 samples, where 9,900 were used for training, and 100 test samples were used to simulate measurements for reconstruction. The images were resized to 256 × 256 and normalized to [−1, 1], following standard diffusion model practices to ensure stable training. The same dataset, preprocessing, and split were maintained for all models. The models were also experimentally evaluated on anatomical images, utilizing the images under the BI-RADS framework24. For evaluation, the test and the anatomical images were used to generate corresponding photoacoustic measurements through a consistent simulation framework, ensuring uniform data generation across all experiments. The images used to simulate the PAT measurements were treated as ground truth, and the reconstructed PAT images were compared against them for quantitative and qualitative assessment. Simulation setup The PAT simulation considers a two-dimensional imaging setup with a circular detector array surrounding the imaging region. A total of 100 transducers are uniformly distributed on a ring of radius 0.11 m, providing full angular coverage. The imaging domain is discretized into a uniform grid with pixel size m, defining the reconstruction field of view. Temporal sampling is performed with a sampling frequency of 2 MHz, and a homogeneous speed of sound of 1500 m/s is assumed. The acquisition window is chosen to capture signals corresponding to the minimum and maximum propagation distances between the image domain and the detector ring. The forward operator is constructed using the curve-driven model-matrix inversion (CDMMI) formulation25, which defines the linear mapping between the initial pressure distribution and the measured signals. For each detector and time sample, acoustic travel distances from all image pixels are computed and converted to arrival times, with contributions assigned based on geometric weighting. This process forms the system matrix A, where rows correspond to detector–time measurements and columns to image pixels. PAT measurements reflect the physical acquisition process represented as a 2D array where rows correspond to individual transducers and columns correspond to discrete time samples. Limited-view conditions are simulated by retaining only a subset of detectors from the full-ring geometry and removing the rest. To systematically assess robustness under incomplete data, synthetic photoacoustic measurements were generated using varying acquisition with detector coverage reduced to 70%, 50%, and 30% of the full-view configuration. Reconstructions from each setting were analyzed to evaluate image quality and structural fidelity under increasing data incompleteness. Implementation details The experimental setup follows the standard implementation protocols as reported in the original works of diffusion-based generative models22. To ensure fair comparison, all the parameters and architectures were closely aligned for all the experiments. Both DDPM/DDIM and score-based models adopt a U-Net–style encoder–decoder backbone with timestep conditioning and skip connections, differing mainly in their residual block design.The DDPM/DDIM models use standard ResNet blocks, while the score-based model follows an NCSN++-style design. The network depth was kept consistent (4 levels), with comparable parameter counts. The DDPM/DDIM framework was trained with 1000 diffusion steps and optimized using Adam. It employed a linear beta schedule ( to ) with a learning rate of . While the score-based model used a VP-SDE (variance preserving stochastic differential equation) formulation with parameters adapted to continuous-time training and a learning rate of During training, noise was injected using a linear variance schedule, ensuring a gradual and stable corruption of the signal. This choice maintains a balance between preserving measurement-consistent information and enabling effective learning of priors, which is particularly important for ill-posed PAT reconstruction. DDPM/DDIM and score-based models use closely related diffusion processes but differ in their mathematical formulation. DDPM model employs a discrete-time diffusion schedule, whereas the score-based model uses a continuous-time variance-preserving SDE. Both approaches model gradual noise corruption and denoising. All experiments were carried out in PyTorch 2.2.1 with CUDA 12.1 and cuDNN 8.9.2, ensuring compatibility with GPU-accelerated training pipelines. The models were trained and evaluated on a high-performance computing (HPC) cluster, using a dedicated GPU node equipped with an NVIDIA Quadro RTX 8000, which provided sufficient memory and computing capability for diffusion-based training. All experiments were conducted in a Python 3.8.5 environment to maintain stability and reproducibility across runs. Inference During inference, DDPM-based reconstruction was performed by iteratively alternating between learned denoising and physics-based measurement consistency updates over 1,000 diffusion steps. At each step, the denoiser generated an intermediate estimate guided by the learned prior, which was subsequently projected onto the measurement space to enforce consistency with the observed photoacoustic pressure data. A similar strategy was employed for DDIM, using only 50 sampling steps to enable faster inference. Score-based models formulate diffusion in continuous time via stochastic differential equations and incorporate measurement consistency through a predictor–corrector scheme that alternates between score-driven denoising and Langevin refinement. This multiscale enforcement allows measurement information to influence the reconstruction through 500 time steps. The implemented DDPM/DDIM reconstruction framework can thus be viewed as an analogue of predictor–corrector sampling. Evaluation metrics The performance of the reconstruction was evaluated using metrics like Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM)26, and Learned Perceptual Image Patch Similarity (LPIPS)27, which together capture complementary aspects of image quality. PSNR measures pixel-wise fidelity and sensitivity to global intensity errors, serving as a baseline indicator of reconstruction accuracy. It measures pixel-wise agreement between a reconstructed image and the reference by relating the maximum signal intensity to the mean squared error. Given a reference image and a reconstructed image , the PSNR is defined as where L denotes the maximum possible pixel value. While PSNR captures pixel-level errors, SSIM evaluates structural similarity by comparing local luminance, contrast, and edge information, making it particularly relevant for assessing anatomical preservation. The SSIM between and is defined as where and are local means, and are local variances, is the local covariance, and are constants for numerical stability. To capture perceptual quality beyond pixel-level similarity, LPIPS is used which measures deep feature representations and is effective at detecting visually implausible textures and artifacts introduced by learning-based methods. A lower LPIPS score indicates that the images are more similar perceptually. where denotes the feature activations at layer l of a pretrained network, are learned channel-wise weights, and and denote the spatial dimensions of the feature maps. Together, these metrics enable a quantitatively meaningful evaluation of diffusion-based reconstructions for ill-posed photoacoustic imaging. Results This section presents a quantitative and qualitative evaluation of DDPM, DDIM, and score-based reconstruction approaches for PAT across a range of acquisition conditions, using both synthetic and anatomical datasets. For DDPM and DDIM, reconstruction performance is further examined with and without the inclusion of the refinement step (corresponding to the last stage in algorithm 1) to assess its impact on image quality and robustness. As shown in Table 1, under full-view (FV) conditions, both models achieved high reconstruction fidelity, with PSNR values around 36 dB and SSIM values above 0.87. The metrics, computed over 100 samples, remain consistent, with DDPM showing a standard deviation of 2.25 in PSNR and 0.016 in SSIM. Similarly, DDIM exhibits comparable variability, with standard deviations of 2.28 for PSNR and 0.015 for SSIM. DDIM consistently outperformed DDPM, yielding slightly higher PSNR and SSIM and lower LPIPS across all coverage settings. As the transducer coverage decreased, reconstruction quality degraded for both models. However, the degradation was substantially less severe when refinement was applied. In particular, under limited-view (LV-50%) and limited-view (LV-30%) settings, the refinement step in the algorithm improved PSNR significantly and reduced perceptual error (LPIPS). Notably, DDIM with refinement maintained structural similarity even under highly limited-views, whereas unrefined reconstructions exhibited more structural collapse. Representative reconstructions obtained with and without refinement are shown in Figures 1 and 2 for DDPM and DDIM, respectively. Without refinement, reconstructions under limited coverage exhibit strong artifacts and incomplete vascular continuity. These artifacts become increasingly dominant as the coverage decreases. The introduction of the refinement step mitigates these effects by preserving fine vascular branches and suppressing background noise, particularly in limited-view (LV-50%) cases. These visual improvements are consistent with the quantitative metric observed in Table 1 and indicate that refinement plays a critical role in stabilizing diffusion-based reconstructions under ill-posed acquisition conditions. Table 1Reconstruction performance of DDPM and DDIM for 100 synthetic test samples under full-view and limited-view settings, reported with and without the refinement step.SettingMetricDDPMDDIMFVLV-70LV-50LV-30FVLV-70LV-50LV-30Without refinementPSNR36.3926.5615.798.8736.7027.0417.6810.31SSIM0.8750.4500.2290.0730.8980.4870.2650.092LPIPS0.0210.2220.4360.6180.0160.1950.3920.563With refinementPSNR36.9534.2729.0817.2937.1034.5729.7416.05SSIM0.9250.8280.5750.2560.9350.8490.6120.268LPIPS0.0140.1270.2590.4640.0120.1140.2370.417 Reconstruction performance of DDPM and DDIM for 100 synthetic test samples under full-view and limited-view settings, reported with and without the refinement step. DDPM reconstructions shown for full-view and limited-view (70%, 50%, and 30%) acquisition, with and without the refinement step. Ground-truth images are displayed in the left column for reference. DDIM reconstructions shown for full-view and limited-view (70%, 50%, and 30%) acquisition, with and without the refinement step. Ground-truth images are displayed in the left column for reference. The score-based model was evaluated separately, and its quantitative performance is summarized in Table 2. Under full-view and moderately limited-view (LV-70%), the score-based model achieved high reconstruction fidelity, with PSNR of around 46 dB and extremely low LPIPS values (0.0006), indicating strong perceptual fidelity. The results are also highly consistent across 100 samples, with standard deviations of 0.89 for PSNR and 0.0018 for SSIM. However, under severe view limitation (LV-30%), a marked degradation in PSNR was observed dropping to 23.75 dB and LPIPS increasing sharply. These results suggest that while the score-based model is highly effective when sufficient measurement information is available, its performance deteriorates more rapidly than DDPM and DDIM based approaches as the inverse problem becomes increasingly ill-posed. Table 2Reconstruction performance of Score- based model reconstruction for 100 synthetic test samples under full-view and limited-view settings.MetricFVLV-70LV-50LV-30PSNR46.7545.7340.8723.75SSIM0.9760.9880.98890.815LPIPS0.00060.00070.00190.077 Reconstruction performance of Score- based model reconstruction for 100 synthetic test samples under full-view and limited-view settings. To further examine model behavior, sampling trajectories at selected diffusion steps were visualized (Fig. 3). DDPM exhibited a gradual and stochastic denoising process across 1000 diffusion steps, with early ring-like patterns reflecting the circular wavefront nature of photoacoustic acquisition. These geometry-induced artifacts were progressively suppressed as denoising proceeded, allowing vascular structures to emerge over time. This process requires approximately 41.603 seconds per image. In contrast, DDIM converged substantially faster, with stable vascular features appearing early in the sampling process and remaining consistent across 50 diffusion steps, leading to a reconstruction time of about 2.076 seconds per image. The score-based model showed a smooth evolution from noise-dominated states to coherent reconstructions over 500 diffusion steps and requires 40.187 seconds per image, with measurement information continuously incorporated to preserve data fidelity. Sampling path visualization at selected diffusion steps for DDPM, DDIM, and score-based models with the ground truth shown for reference. Performance on anatomical data was assessed by evaluating all models on representative anatomical images under full-view and limited-view (50%) conditions (Fig. 4). Under full-view acquisition, all methods produced visually plausible reconstructions consistent with the quantitative metrics reported in Table 3. Under limited-view conditions, a clear degradation in reconstruction quality was observed across all models, as reflected by corresponding decreases in quantitative performance. DDPM and DDIM reconstructions retained overall structure but showed increased streaking artifacts. Score-based reconstructions often displayed distorted or exaggerated vascular features, suggesting increased reliance on the learned prior when measurement information was insufficient. While the score-based model achieved the highest PSNR and lowest LPIPS under full-view conditions, its SSIM decreased substantially under limited-view acquisition, indicating reduced structural reliability. To provide practical context for diffusion-based methods, two conventional baselines, Tikhonov and Split Bregman Total Variation (SBTV), were included for comparison. Tikhonov enforces smoothness, while SBTV yields sharper, more structured reconstructions. Under full-view, SBTV achieves higher PSNR and SSIM. In limited-view settings, both degrade, but SBTV performs slightly better, whereas Tikhonov loses detail. Overall, diffusion-based models consistently outperform these conventional approaches in both settings. Anatomical reconstructions for full-view and limited-view (50%) acquisition are shown for DDPM, DDIM, score-based, Tikhonov and SBTV, alongside the corresponding ground truth. Table 3Anatomical scan reconstruction metrics of DDPM, DDIM, score-based models, Tikhonov and SBTV evaluated on five anatomical cases under full-view and limited-view (50%) acquisition. Bold values indicate the best metric among all the models, across individual anatomical cases.ModelMetricFull-viewLimited-view (50%)abcdeabcdeDDPMPSNR30.0932.0531.4432.2030.4328.2328.2422.7827.3821.30SSIM0.9710.8830.9860.9750.9690.8450.6650.8360.7780.592LPIPS0.0230.0180.0390.0190.0260.2360.3220.2420.2780.410DDIMPSNR30.4332.0731.5432.7030.4728.3125.6323.5526.3422.21SSIM0.9720.8940.9860.9760.9690.8760.6820.8550.7750.644LPIPS0.0220.0180.0390.0180.0260.1970.3130.2480.2510.378Score-basedPSNR35.3638.7237.2729.2232.6519.5127.6520.2623.3224.07SSIM0.9390.9630.9900.9230.9500.3520.8180.6910.6810.720LPIPS0.0370.0090.0020.0100.0030.7350.1760.01890.0970.049SBTVPSNR6.5112.5919.2118.7217.798.6210.1014.8214.5813.92SSIM0.1500.1180.9230.7210.5950.3140.1350.7440.5140.423LPIPS0.4410.5550.1150.1470.3010.4910.5870.3450.3080.417TikhonovPSNR6.3713.029.5912.713.765.7411.948.8911.8312.43SSIM0.1050.2160.1990.1340.1360.0480.0780.1020.0810.067LPIPS0.3410.3520.3760.2640.3050.4270.4690.4240.3560.395 Anatomical scan reconstruction metrics of DDPM, DDIM, score-based models, Tikhonov and SBTV evaluated on five anatomical cases under full-view and limited-view (50%) acquisition. Bold values indicate the best metric among all the models, across individual anatomical cases. Discussion This study presents a comparative evaluation of diffusion-based reconstruction methods for PAT under varying acquisition conditions, highlighting their behaviour to data incompleteness. A consistent trend is observed among probabilistic (DDPM), deterministic (DDIM), and score-based approaches for both synthetic and anatomical data. For DDPM and DDIM, incorporating a refinement step improves stability under limited-view conditions by enforcing measurement consistency. DDPM relies on a slow, stochastic denoising process, and even with 1000 steps, struggles to recover missing measurements. However, the refinement step helps mitigate this by enforcing measurement consistency and preserving structure. DDIM achieves improved stability with significantly fewer timesteps due to its deterministic sampling, which reduces stochastic perturbations and limits exposure to highly noisy intermediate states. Analysis of sampling trajectories further highlights these differences. However, DDPM remains effective when sufficient data constraints are present, as its multi-step denoising enables gradual refinement of image structures and supports high-fidelity reconstruction and remains foundational model. It continues to serve as a foundational reference for subsequent variants. The score-based model demonstrates strong performance when measurement information is sufficiently informative, achieving better performance metrics under full-view conditions. However, under severe view limitations, it exhibits structural distortions, suggesting increased reliance on the learned prior as measurement constraints weaken. This reflects the need to balance prior strength with data fidelity in ill-posed settings19. The smooth evolution of score-based sampling indicates effective integration of measurement information across noise scales. Experiments on unseen anatomical data support the overall findings and demonstrate generalization beyond the training distribution. While all methods produce plausible reconstructions under full-view acquisition, limited-view settings introduce streaking artifacts. The observed performance differences among the models can be attributed to how measurement information is incorporated during the diffusion process. DDPM and DDIM operate through discrete denoising steps with projection-based data consistency under a fixed noise schedule, where early high-noise stages can weaken measurement influence, making reconstruction quality dependent on noise scheduling and sampling strategy28,29. In contrast, score-based models adopt a continuous-time formulation with a predictor–corrector framework, that enforces measurement consistency across noise scales. This multiscale conditioning mitigates early loss of data fidelity and reducing over-reliance on the learned prior16,17. These observations highlight potential advantages of continuous, scale-consistent conditioning, while also indicating discrete diffusion methods may require further refinement for severely ill-posed inverse problems30. These trends are consistent with the improvements observed in our experiments when an explicit refinement step is incorporated into the sampling algorithm. Limitations The findings of this study should be interpreted in light of some limitations. The study considered a limited set of diffusion formulations and reconstruction strategies, and alternative noise schedules may lead to different outcomes. Prior work by Nichol et al. showed that linear noise schedules destroy signal information more rapidly than cosine schedules, motivating improved variance schedules for better sample quality31. Adopting smoother or adaptive noise schedules that better preserve information at high noise levels could therefore improve robustness in ill-posed photoacoustic reconstruction settings. The refinement step was implemented using a gradient descent approach, and other refinement strategies could further influence stability and reconstruction performance. Despite these limitations, the three models considered in this work represent foundational diffusion frameworks upon which many recent and more advanced diffusion-based reconstruction methods are built, making them well suited for a systematic comparative analysis. Conclusion This study examined diffusion-based reconstruction approaches for photoacoustic tomography, focusing on probabilistic (DDPM), deterministic (DDIM), and score-based formulations under different acquisition conditions. By evaluating both synthetic data and anatomical scans, the analysis shows that reconstruction performance is influenced by how diffusion dynamics and measurement consistency are incorporated, especially when data coverage is limited. Experiments incorporating an explicit refinement step showed improved stability and reconstruction quality for both DDPM and DDIM in ill-posed settings. The refinement step in the algorithm reduced artifacts and helped maintain structural continuity as view coverage decreased. Between the two diffusion models, DDIM consistently reached stable reconstructions using far fewer sampling steps, resulting in improved efficiency without compromising reconstruction quality. While score-based models demonstrated more robust reconstruction quality across acquisition settings, the algorithm converges with more diffusion time steps compared to DDIM, which may limit inference efficiency in time-critical applications. Overall, the results indicate that successful reconstruction depends on the interaction between the diffusion process and physics-based constraints. Both the choice of diffusion formulation and the way noise is introduced and removed during sampling play a critical role in determining robustness under limited-view conditions. Deterministic diffusion combined with refinement appears well-suited for scenarios with incomplete data, while score-based approaches are most effective when the measurement setting is well-conditioned. These findings provide practical guidance for adapting diffusion-based reconstruction methods to photoacoustic tomography applications. Supplementary Information Supplementary Information. Author contributions Shibili Said and Imad Barhumi contributed equally to this work. Data availability The dataset used in this study is publicly available and can be accessed from the corresponding repositories. The synthetic dataset used in this study is available on GitHub at https://github.com/sreemanti-dey/diffusion_for_PAT. The clinical scans are available at https://radiopaedia.org/articles/breast-imaging-reporting-and-data-system-bi-rads-2 [Case 2 (BI-RADS 4), Case 3 (BI-RADS 5), Case 5 (BI-RADS 5), and Case 9 (BI-RADS 4)]. Code availability The code used to implement the algorithm and generate the results presented in this study is publicly available on GitHub at https://github.com/shibilisaid/DM_COMPARISONFOR_PAT. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Information The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57923-2. References Associated Data Data Availability Statement The dataset used in this study is publicly available and can be accessed from the corresponding repositories. The synthetic dataset used in this study is available on GitHub at https://github.com/sreemanti-dey/diffusion_for_PAT. The clinical scans are available at https://radiopaedia.org/articles/breast-imaging-reporting-and-data-system-bi-rads-2 [Case 2 (BI-RADS 4), Case 3 (BI-RADS 5), Case 5 (BI-RADS 5), and Case 9 (BI-RADS 4)]. The code used to implement the algorithm and generate the results presented in this study is publicly available on GitHub at https://github.com/shibilisaid/DM_COMPARISONFOR_PAT.