 Original Article
 Open access
 Published:
Noise suppression in photoncounting computed tomography using unsupervised Poisson flow generative models
Visual Computing for Industry, Biomedicine, and Art volume 7, Article number: 24 (2024)
Abstract
Deep learning (DL) has proven to be important for computed tomography (CT) image denoising. However, such models are usually trained under supervision, requiring paired data that may be difficult to obtain in practice. Diffusion models offer unsupervised means of solving a wide range of inverse problems via posterior sampling. In particular, using the estimated unconditional score function of the prior distribution, obtained via unsupervised learning, one can sample from the desired posterior via hijacking and regularization. However, due to the iterative solvers used, the number of function evaluations (NFE) required may be orders of magnitudes larger than for singlestep samplers. In this paper, we present a novel image denoising technique for photoncounting CT by extending the unsupervised approach to inverse problem solving to the case of Poisson flow generative models (PFGM)++. By hijacking and regularizing the sampling process we obtain a singlestep sampler, that is NFE = 1. Our proposed method incorporates posterior sampling using diffusion models as a special case. We demonstrate that the added robustness afforded by the PFGM++ framework yields significant performance gains. Our results indicate competitive performance compared to popular supervised, including stateoftheart diffusionstyle models with NFE = 1 (consistency models), unsupervised, and nonDLbased image denoising techniques, on clinical lowdose CT data and clinical images from a prototype photoncounting CT system developed by GE HealthCare.
Introduction
Xray computed tomography (CT) is a medical imaging modality used for the diagnosis and treatment planning of a wide range of diseases, including stroke, cancer, and cardiovascular diseases. However, because of the potential risks posed by even low doses of ionizing radiation, considerable efforts have been made to enable high diagnostic quality while maintaining a dose as low as reasonably achievable [1, 2]. Photoncounting CT (PCCT), which is based on the latest generation of CT detector technology, can reduce the dose via photon energy weighting and eliminate the effects of electronic noise. PCCT can also enable imaging with a higher spatial resolution and produce singleexposure energyresolved images [3,4,5,6]. However, obtaining high resolution in either space or energy decreases the number of photons in each voxel or energy bin, which increases the image noise. Thus, an excellent noise performance is required, possibly exceeding the capabilities of today’s stateoftheart denoising methods.
In recent years, deep learning (DL) methods have demonstrated remarkable success in lowdose CT (LDCT) and PCCT image denoising [7,8,9,10,11,12,13,14,15,16,17,18,19]. However, these methods are commonly trained using supervised learning, which requires paired data, which are not always available in practice. In particular, paired and perfectly registered clinical images are difficult to obtain, and methods for generating paired data based on simulating lowdose scanning [20] or adding noise maps from phantom scans [21] may be sensitive to imperfect modeling of the systems or mismatches between patient and phantom geometries. In PCCT, pulse pileup affects high and lowdose scans differently, further confounding the training schemes. Therefore, unsupervised and selfsupervised methods are becoming increasingly common [12, 13, 15,16,17].
Diffusionstyle models, such as diffusion and Poisson flow models, have demonstrated considerable success in unconditional [22,23,24,25,26,27,28,29] and conditional image generation [25, 30,31,32,33]. Notably, for medical imaging, this family of models lends themselves very well to inverse problem solving via posterior sampling and has already been demonstrated on a range of inverse problems [17,18,19, 34, 35]. In particular, in the case of diffusion models, it is possible to manipulate the sampling process and retask the network, trained in an unsupervised manner, for inverse problemsolving [17, 31, 34, 35]. This is typically performed in two steps. First, information from the prior distribution, that is, the ground truth data, is obtained by estimating its timedependent score function via denoising score matching, exactly as when training an unconditional generator. Once equipped with the estimated score function, samples can be drawn from the desired posterior distribution by augmenting the sampling process with a data consistency step that regularizes the generative process and forces the sample to be consistent with the input (conditioning) image. Moreover, running the sampling process using the initial sample from a prior noise distribution is unnecessary. Indeed, it is even beneficial to ‘hijack’ the sampling process by inserting a version of the condition image at some stage of the reverse diffusion [31, 35]. Hijacking will not only help regularize the problem further but will also result in faster sampling; a smaller number of function evaluations (NFE) are required to achieve the desired image quality. Note that this is an efficient strategy because the learned score function can be successfully used to solve a range of different inverse problems or downstream tasks without retraining [34, 35]. It is also possible to solve the inverse problem directly using supervised learning [18, 19]. However, the supervised approach requires paired data.
The NFE required for diffusionstyle models, such as diffusion and Poisson flow models, may be of the order 10^{1}–10^{3} for both conditional (imagetoimage) and unconditional (noisetoimage) generation. This limits their use in applications where speed is critical, such as clinical CT image denoising. Efforts to reduce the required NFE include the use of efficient ordinary differential equation (ODE) samplers [26] and distillation techniques [36]. Consistency models [37] build on the probability flow ODE formulation of diffusion models and learn a consistency network that maps any point on the trajectory to its initial point, including the final point, which is a sample from the prior noise distribution. Thus, it achieves singlestep sampling (NFE = 1). A consistency model may be distilled from a pretrained diffusion model using socalled consistency distillation (CD) or obtained as a standalone model using consistency training.
This study presents a novel image denoising technique for lowdose and photoncounting CT that extends posterior sampling Poisson flow generative models (PPFM) [19] to cases where paired data are not available. The PPFM is a diffusionstyle posterior sampling image denoising method that exploits the added robustness of the PFGM++ framework to enable NFE = 1 while maintaining high image quality. In particular, an unconditional PFGM++ was first trained for image generation on randomly extracted patches from the prior, ground truth, and images. The sampling process was subsequently hijacked and regularized, as in PPFM, to ensure consistency with the input, condition, and image. The main contributions of this study are as follows. (1) Unsupervised PPFM, a novel imagedenoising technique that extends PPFM [19] to the case when paired data are unavailable, is presented. (2) It is demonstrated that it is possible to efficiently train the network in an unsupervised manner on patches extracted from ground truth data and subsequently manipulate the sampling process to denoise fullresolution images. Training on randomly extracted patches is more efficient in terms of graphics memory requirements and provides additional regularization. (3) The proposed method contains a posterior sampling diffusion model (EDM [27]) as a special case when D → ∞. The results indicate that the added flexibility of choosing D as a hyperparameter results in improved performance when D → ∞ as is the case for diffusion models. (4) The proposed method was evaluated using clinical lowdose CT (LDCT) images and clinical images from a prototype PCCT system developed by GE Healthcare (Waukesha, WI, USA) [38]. It is demonstrated that the proposed method performs competitively compared with current stateoftheart diffusionstyle models with NFE = 1 and consistency models [37]. Notably, the consistency model was trained in a supervised manner, whereas the proposed method was unsupervised. Despite imposing a significantly laxer data requirement, the proposed method performs both quantitatively and qualitatively. In addition to consistency models, the proposed method was compared with popular supervised, unsupervised, and nonDLbased image denoising techniques.
The code used in this study is available at: https://github.com/dennishein/pfgmpp_PCCT_denoising.
Methods
Diffusion models and PFGM++
Diffusion models (EDM [27]) and PFGM++ [29] work by iteratively denoising images by following a physically meaningful trajectory inspired by nonequilibrium thermodynamics and electrostatics, respectively. Despite the widely different underlying physics, diffusion models and PFGM++ are intimately connected in theory and practice. The training and sampling processes of PFGM++ converge to those of the diffusion models in the D → ∞, \(r=\sigma \sqrt{D}\) limit [29]. Thus, diffusion models were incorporated as special cases of PFGM++. In addition, it is possible to reuse the training and sampling algorithms in the EDM [27] for PFGM++ using a simple change in variables and an updated prior noise distribution [29]. Expanding on the probability flow ODE formulation in ref. [25], the EDM [25] writes the ODE as
where \({\nabla }_{x}log{p}_{\sigma (t)}({\varvec{x}})\) is the timedependent score function of the perturbed distribution and σ(t) is the noise scale. This ODE defines the trajectory between an easytosample prior noise distribution, a simple Gaussian distribution, and the data distribution of interest. Intuitively, running the ODE forward or backward in time nudges the sample toward or away from the prior noise distribution. Notably, Eq. 1 depends only on the data distribution via the gradient of the loglikelihood, also known as the score function. Let p(y) represents the data distribution, p(σ) represents the distribution of noise scales, and \({p}_{\sigma }({\varvec{x}}{\varvec{y}})=\mathcal{N}\left({\varvec{y}}, {\sigma }^{2}{\varvec{I}}\right)\) the Gaussian perturbation kernel. The score function can then be estimated using a perturbationbased objective:
where λ(σ) denotes a weighting function. Given the estimation of the timedependent score function, a sample was generated by solving Eq. 1, using an iterative ODE solver, starting from an initial sample from the prior noise distribution.
PFGM++ [29] operates by treating Ndimensional data as electric charges in an (N + D)dimensional augmented space. Let \(\widetilde{{\varvec{y}}}:=\left({\varvec{y}}, 0\right)\in {\mathbb{R}}^{N+D}\) and \(\widetilde{{\varvec{x}}}:=\left({\varvec{x}},\boldsymbol{ }{\varvec{z}}\right)\in {\mathbb{R}}^{N+D}\) denote the augmented ground truth and perturbed data, respectively. The object of interest in highdimensional electric fields is
where S_{N+D−1}(1) is the surface area of the unit (N + D − 1)sphere and p(y) is the ground truth and data distribution. However, owing to the rotational symmetry of the Ddimensional cylinder,\({\sum }_{i=1}^{D}{z}_{i}^{2}={r}^{2},\forall r>0\), dimensionality reduction is possible [29]. In fact, it suffices to track the norm of the augmented variables \(r=r\left(\widetilde{{\varvec{x}}}\right):={\Vert z\Vert }_{2}\). For notational brevity, \(\widetilde{{\varvec{y}}}:=\left({\varvec{y}}, 0\right)\in {\mathbb{R}}^{N+1}\) and \(\widetilde{{\varvec{x}}}:=\left({\varvec{x}},\boldsymbol{ }\mathbf{r}\right)\in {\mathbb{R}}^{N+1}\) are redefined.
The ODE of interest is:
where \({{\varvec{E}}\left(\widetilde{{\varvec{x}}}\right)}_{{\varvec{x}}}\), and \({{\varvec{E}}\left(\widetilde{{\varvec{x}}}\right)}_{r}\), a scalar, denote the x and r components of \({\varvec{E}}\left(\widetilde{{\varvec{x}}}\right)\), respectively. Equation 3 defines a surjection between the data on the r = 0 hyperplane and an easytosample prior noise distribution on the r = r_{max} hypercylinder [29]. For the diffusion models, PFGM++ employs a perturbationbased objective. Let p(r) be the training distribution over r and \({p}_{r}({\varvec{x}}{\varvec{y}})\) be the perturbation kernel. Then, the objective of interest is
Now, if \({p}_{r}\left({\varvec{x}}{\varvec{y}}\right)\propto 1/{\left({\Vert {\varvec{x}}{\varvec{y}}\Vert }_{2}^{2}+{r}^{2}\right)}^{\frac{N+D}{2}}\) then it is possible to show that the minimizer of Eq. 4 is \({f}_{\theta }^{*}\left(\widetilde{{\varvec{x}}}\right)=\sqrt{D}{{\varvec{E}}\left(\widetilde{{\varvec{x}}}\right)}_{{\varvec{x}}}\cdot {E\left(\widetilde{{\varvec{x}}}\right)}_{r}^{1}\). As for diffusion models, a sample can be generated by solving \(dx/dr={{\varvec{E}}\left(\widetilde{{\varvec{x}}}\right)}_{{\varvec{x}}}/{E\left(\widetilde{{\varvec{x}}}\right)}_{r}={f}_{\theta }^{*}\left(\widetilde{{\varvec{x}}}\right)/\sqrt{D}\), starting from an initial sample from the prior noise distribution, \({p}_{{r}_{max}}\), using an iterative ODE solver.
Problem formulation
The problem of obtaining a high quality reconstruction \(\widetilde{{\varvec{y}}}\in {\mathbb{R}}^{N}\) of \({\varvec{y}}\in {\mathbb{R}}^{N}\) based on noisy observations \(\varvec{c}=\mathcal{F}\left({\varvec{y}}\right)\in {\mathbb{R}}^{N}\) is treated as a statistical inverse problem, where \(\mathcal{F}:{\mathbb{R}}^{N}\to {\mathbb{R}}^{N}\) is a ‘catchall’ noise degradation operator, including factors such as quantum noise [7], and N: = n × n. Thus, it is assumed that the data follow a prior distribution \(y\sim p\left(y\right)\) and the objective is to enable sampling from the posterior \(p\left({\varvec{y}}{\varvec{c}}\right)\). This approach to inverse problemsolving is referred to as posterior sampling. In this study, y is treated as “ground truth” despite the fact that it may contain noise and artifacts.
Image denoising via posterior sampling
The proposed method has two main components: (1) A learned PFGM++ trained in an unsupervised manner for unconditional image generation. (2) A sampling scheme that regularizes the generative process and enforces consistency with the input, conditions, and images. By combining the information from the prior distribution, p(y), embedded in the learned empirical electric, the field can be sampled with a modified sampling scheme to ensure data consistency, and sampled from the desired posterior p(yc), thus providing a solution to the inverse problem. This strategy has been successfully applied to diffusion models for a range of inverse problems [17, 31, 34, 35]. The key idea of this study was to extend this approach to the case of PFGM++. Direct extension to PFGM++ is feasible because of its intimate connection with diffusion models. The theoretical analysis is left for future work and empirical support is provided for obtaining a sample from the desired posterior. In particular, it is demonstrated that \(\widehat{{\varvec{y}} }\approx {\varvec{y}}\). Similar to PFGM++ [29], the proposed sampling algorithm is reused from ref. [27] with an updated prior noise distribution. This is feasible because of the hyperparameter translation formula, \(r = \sigma \sqrt{D}\), \(\widetilde{{\varvec{x}}}:=\left({\varvec{x}}, r\right)\), the fact that \(\sigma \left(t\right)=t\) in ref. [27], and by a change of variable \(d{\varvec{x}}={f}_{\theta }^{*}\left(\widetilde{{\varvec{x}}}\right)/\sqrt{D}dr={f}_{\theta }^{*}\left(\widetilde{{\varvec{x}}}\right)dt\) since \(dr=d\sigma \sqrt{D}=dt\sqrt{D}\). The proposed sampling algorithm is detailed in Algorithm 1 with updates to sampling vs [29] highlighted in blue. The first update is hijacking, and instead of starting from an initial sample for the prior noise distribution, the reverse process is hijacked at some \(i = \tau \in {\mathbb{Z}}_{+}\), τ < T and simply inject the condition image x_{τ} = c. In addition to hijacking, data consistency, regularization, and steps are considered. It must be updated for various inverse problems. For image denoising, simply using an identity map is sufficient. Thus, x_{i+1} is simply mixed with x_{τ} = c, where c is the condition image with weight w ∈ [0, 1]. Note that the proposed sampling algorithm is identical to that used in ref. [19] except that the network has now been trained using unsupervised learning and thus does not take the condition image c as an additional input.
Experiments
Datasets
Mayo LDCT data: For training and validation, a publicly available dataset from the Mayo Clinic used in the American Association of Physicists in Medicine (AAPM) Lowdose CT Grand Challenge [39] was used. This dataset contains images from ten patients reconstructed using medium (D30) and sharp (D45) kernels. Images were also available for 1 mm and 3 mm slice thicknesses. All the images had a matrix size of 512 × 512. In this study, a slice thickness of 1 mm and B30 reconstruction kernel were used. The first eight patients were used for training, yielding 4800 slices, and the last two for validation, yielding 1136 slices. Although paired normaldose CT (NDCT) and LDCT images were available, only NDCT images were used for training the proposed method.
PCCT data: As test data, the images of two patients from a clinical study (Swedish Ethics Review Agency 2020–04638 and 2021–01092 and prospectively consented IRB review UWIRB: 2022–1043) of a prototype silicon detectorbased photoncounting system developed by GE Healthcare [38] were used. The patients were scanned at the Karolinska Institutet, Stockholm, Sweden (Case 1, effective diameter 28 cm, CTDI_{vol} = 10.12 mGy) and at the University of WisconsinMadison, Madison, WI (Case 2, effective diameter 36 cm, CTDI_{vol} = 27.64 mGy) using scan parameters listed in Table 1. Virtual monoenergetic images of 70 keV were reconstructed using filtered backprojection on a 512 × 512pixel grid with a 0.42 mm slice thickness.
Implementation details
Each network, D ∈ {64, 128, 2048} and D → ∞, was trained with a batch size of 32 using Adam [40] and a learning rate of 2 × 10^{−4} for 105 iterations. The DDPM++ architecture was used with a channel multiplier of 128 channels per resolution [1, 1, 2, 2, 2, 2, 2] and selfattention layers at resolutions of 16, 8, and 4. The preconditioning, exponential moving average schedule, and nonleaky augmentation suggested in ref. [27] with an augmentation probability of 15% were employed. Additionally, to prevent further overfitting, the dropout probability was set to 10%. To ensure efficient training, the network was trained on randomly extracted 256 × 256 patches. In other words, the unconditional generator was not trained to generate fullresolution CT images but rather patches extracted from CT images. In addition to facilitating efficient training, this served to further augment the dataset and, therefore, helped prevent overfitting. To reduce the graphics memory requirements, mixed precision was used in the training. Note that the same configuration, but with 512 × 512 images, would exceed the memory available on an NVIDIA A6000 48 GB GPU. To achieve NFE = 1, τ: = T − 1, was set. T and w, which are crucial hyperparameters in Algorithm 1, were set by grid search over T ∈ {4, 8, 16, 32, 64} and w ∈ {0.5, 0.6, 0.7, 0.8, 0.9, 1.0} using Learned Perceptual Image Patch Similarity (LPIPS) [41] on the validation was set as the selection criteria. The lowest (best) LPIPS was achieved for T = 8 and w = 0.5. Because w = 0.5 is a corner case, was also tried w = 0.4. However, this did not improve the performance. The proposed method is referred to as ‘singlestep’ despite including a second step, regularization, since the time required from this operation is negligible and NFE = 1.
Comparison to other methods
The results were compared with those of popular nonDLbased, supervised, and unsupervised imagedenoising techniques. For nonDLbased image denoising, a version of BM3D [42] was chosen because it has been shown to be the best performing method in the nonDL category for LDCT image denoising [7]. Bm3d.py (https://pypi.org/project/bm3d/) was used, and the parameter σ_{BM3D} was set by measuring the standard deviation in a flat regionofinterest (ROI) in the LDCT validation data. For the supervised techniques, the WGANVGG [9], consistency models [37], and PPFM [19] were used. The WGANVGG was trained on randomly extracted 64 × 64 from the LDCT training data with hyperparameters, as specified in ref. [9]. The consistency models [37] are stateoftheart diffusionstyle models with an NFE of 1. However, consistent models have been developed for generating unconditional images. Notably, ref. [19] was the first study to implement a consistency model for conditional (imagetoimage) generation. It is reasonable to surmise that the ‘trick’ of feeding the condition image as additional input to the network to directly learn a trajectory to the posterior distribution of interest, a technique successfully used for diffusion models [30, 33, 43] and PFGM++ [19], will also work for CD. This hypothesis is empirically supported by ref. [19]. Minimal adjustments were made to the official implementation (https://github.com/openai/consistency_models) to feed conditional images as additional input. The network was trained on randomly extracted 256 × 256 patches from the LDCT data. Random rotations and mirrors were used for data augmentation. The hyperparameters for sampling and training were set as in the LSUN 256 × 256 experiments (https://github.com/openai/consistencymodels/blob/main/scripts/launch.sh), except for the batch size, which had to be reduced to four to fit in the memory of a single NVIDIA A6000 48 GB GPU. An EDM was first trained for 3 × 10^{5} iterations and then distilled into a consistency model for 6 × 10^{5} iterations. This model is referred to as CD. The training and sampling for the PPFM are specified in ref. [19]. The D = 64 case was used because this yields the best performance. The results for BM3D, WGANVGG, CD, and PPFM were derived directly from ref. [19]. Finally, as an example of an unsupervised method, Noise2Void [44] was used. The same hyperparameters as those for the BSD68 dataset in the original paper were used, using code from the official repository (https://github.com/juglab/n2v).
Evaluation methods
For quantitative image quality assessment, the structural similarity index (SSIM) [45], peak signaltonoise ratio (PSNR), and LPIPS [41] were employed. SSIM and PSNR are wellestablished metrics in medical imaging literature, but these relatively simple metrics do not necessarily correspond well to human perception [41]. For instance, the PSNR is inversely proportional to the ℓ_{2} Euclidean distance, and this simple pixelwise metric correlates well with human perception. This was particularly evident in the case of oversmoothing. To overcome some these issues, ref. [41] suggests using pretrained convolutional neural networks (CNNs) as feature extractors, as is the case for perceptual loss functions. The resulting metrics, called LPIPS, were shown to correspond more closely to human perception than more traditional metrics such as SSIM and PSNR in a range of experiments. In this study, the official implementation of LPIPS (https://github.com/richzhang/PerceptualSimilarity) with AlexNet [46] was used as the feature extractor.
In addition to the evaluation with numerical metrics, two radiologists with 6 and 25 years of experience respectively performed a visual assessment of the image quality of the denoised images. CD was used as the reference, and the observers were asked to grade the image quality resulting from each of the six denoising methods relative to the CD image for the LDCT image in Fig. 1 and the PCCT images in Figs. 4 and 6 (in total 18 comparisons). Grading was performed on a fivepoint Likert scale ranging from 2 (significantly worse) to +2 (significantly better) in five categories: noise, contrast, sharpness, artifacts and overall image quality. The unprocessed PCCT image and the regular LDCT and NDCT images are also displayed for reference.
Results
The quantitative results, mean, and standard deviation of the LPIPS [41], SSIM [45], and PSNR for the LDCT validation set are listed in Table 2. Additionally included is the average time, in seconds, to process a single slice. The notebook used from the official implementation of Noise2Void does not utilize GPU acceleration, which should be considered when comparing the evaluation speeds. The overall top performer was PPFM. However, as mentioned above, SSIM and PSNR do not necessarily correspond closely to human perception. As expected, the proposed method is bounded in performance by PPFM because it is unsupervised. Moreover, it can be seen that the proposed method for D = 128 and D = 64 is better in terms of LPIPS than WGANVGG, a supervised method. Compared with CD, a stateoftheart diffusion style model with NFE = 1, the performance of the proposed method was slightly worse. Importantly, CD was trained in a supervised manner, whereas the proposed method was unsupervised. Thus, the proposed method with D = 128 performed competitively. Comparing the two unsupervised methods, Noise2Void and the proposed method, it can be seen that the latter performs favorably. Noise2Void yields only a marginal improvement over LDCT images. The qualitative results, along with LPIPS, SSIM, and PSNR, for a representative slice from the Mayo LDCT validation data are shown in Fig. 1. For brevity, the results for the proposed method are shown only for D → ∞, and the best performance, D = 128. The former is included as an interesting case because it corresponds to a diffusion model instead of PFGM++. This patient had liver metastasis, and magnification of the ROI were included, as shown in Figs. 2a and b for NDCT and LDCT images, respectively, for reference. BM3D, as shown in Fig. 2c, performed well in terms of noise suppression while preserving the salient details. Nevertheless, this comes at the cost of artifacts that make the image appear smudgy. WGANVGG, CD, and PPFM, all supervised methods, as shown in Figs. 2d, e, and f, suppress noise effectively and keep the key details intact. A yellow arrow is overlaid to indicate the detail that appears in CD, shown in Fig. 2e, but in none of the other images, including NDCT and LDCT. It appears that CD has added a feature to the image that appears realistic, but that is not genuine, given that LDCT and NDCT images are used as references. Such inaccurate removal or addition of details is loosely referred to as a hallucination [2]. Noise2Void, shown in Fig. 2g, appears to essentially reproduce the LDCT image. However, quantitatively, there was only a marginal improvement. However, as shown in Fig. 2i, the proposed method (with D = 128) effectively suppressed noise while keeping the salient features intact. Qualitatively, it is difficult to discriminate between the proposed method with D = 128, which is an unsupervised method, and PPFM, which is shown in Fig. 2f, a supervised method. Quantitatively, for this particular slice, it can be seen that PPFM performs slightly better. Comparing D → ∞, shown in Fig. 2h, with D = 128 in Fig. 2i demonstrates the performance gains afforded by the PFGM++ framework. In particular, the proposed method with D → ∞ appeared oversmoothed and somewhat blurred.
The results of the ablation study of the proposed sampler are shown in Fig. 3. Recall the problem formulation, the objective is to obtain an estimate \(\widehat{{\varvec{y}}}\in {\mathbb{R}}^{\rm N}\) of \({\varvec{y}}\in {\mathbb{R}}^{\rm N}\) based on a noisy observation \(c=\mathcal{F}\left({\varvec{y}}\right)\in {\mathbb{R}}^{\rm N}\). By addressing this statistical inverse problem, the solution is a sample from the posterior \(\widehat{{\varvec{y}}}\sim p\left({\varvec{y}}{\varvec{c}}\right)\). Figures 3a and b show the NDCT (y) and LDCT (c) images, respectively, for ease of reference. In Fig. 3c, the regularization (τ = T − 1, w = 1) is hijacked but omitted. This results in an image that has been very aggressively denoised. This is a direct consequence of the small T and corresponding large step size. This represents a further clean demonstration of how SSIM and PSNR fail to adequately penalize blurring, as Fig. 3c appears blurry to a human observer, yet performs very well according to SSIM and PSNR. LPIPS, on the other hand, penalizes this heavily. In Fig. 3d, regularization occurs, but no hijacking (τ = 0, w = 0.5). In this case, an initial sample from a prior noise distribution was first used. The amount of regularization in this setting appeared to be excessive, and an image where \(\widehat{{\varvec{y}} }\approx {\varvec{c}}\) was recovered. Finally, in Fig. 3e, hijacking and regularization (τ = T − 1, w = 0.5) were employed, resulting in a very pleasing image, where \(\widehat{{\varvec{y}} }\approx {\varvec{y}}\). Consequently, there was a significant reduction (improvement) in LPIPS. Hijacking and regularization allows repurposing of the pretrained image generator for the task of image denoising. Conditioning on the LDCT (c) image via hijacking and regularization enforces a sample from the posterior \(\widehat{{\varvec{y}}}\sim p\left({\varvec{y}}{\varvec{c}}\right)\), where \(\widehat{{\varvec{y}} }\approx {\varvec{y}}\) is the normal dose counterpart to c. In other words, a Bayesian inference is performed.
Qualitative results for a representative slice in the first PCCT scan are shown in Fig. 4 with a magnification of the indicated ROI in Fig. 5. Owing to the lack of “ground truth” images, qualitative evaluation is used. It is also difficult to discriminate signals from noise in small, lowcontrast details. Good performance is simply defined as accurately reproducing the unprocessed image, as shown in Figs. 4a and 5a, but with a lower noise level. BM3D, in Figs. 4b and 5b, appears to generalize poorly from LDCT data. This is likely due to the differences in noise characteristics; one would need to reestimate σ_{BM3D}. All other methods seem to generalize well in the sense that there are no major changes in performance. There is a significant performance gain for the proposed method with D = 128 compared to D → ∞. This was most visible in the magnified ROI in Fig. 7 as D → ∞, shown in Figs. 4g and 5g, is significantly more blurry than D = 128, shown in Figs. 4h and 5h. The yellow arrow indicating the detail of interest has been overlaid. Because there is no ground truth and a single slice is considered, it cannot be definitively stated that this is not just a noise spike. However, because it is clearly visible in an unprocessed image, it needs to be visible in the processed images. As can be seen, this is indeed the case with the notable exception of CD, shown in Figs. 4d and 5d. Moreover, the contrast of this detail appears to vary and seems to be much more welldefined for the proposed method with D = 128 than for WGANVGG, as shown in Figs. 4c and 5c. Hence, it appears that the proposed method, despite being unsupervised, can perform competitively, even when compared to supervised methods such as WGANVGG and CD.
The results of a representative slice in the second PCCT scan are shown in Figs. 6 and 7. Again, no “ground truth” is available. BM3D, in Figs. 6b and 7b, appeared to perform better than the first PCCT test case. The denoising performance is now more aligned with that observed for the LDCT validation data. This difference in performance is likely due to the differences in noise characteristics and a lack of generalization. The lack of generalization does not appear to be an issue for any other method, as the performance is consistent across the validation and test data. A yellow arrow is again placed to indicate the details of interest, in this case, the fat in the back muscle. The contrast is difficult to assess qualitatively when comparing the proposed method with D → ∞, shown in Figs. 6g and 7g, and D = 128, shown in Figs. 6h and 7h, owing to the large difference in the noise level. D → ∞ is definitely oversmoothed, and thus blurry; however, the contrast of this particular detail seems to be fairly well preserved. The proposed method with D = 128 again performs very competitively compared with WGANVGG, as shown in Figs. 6c and 7c, as can be seen when considering the contrast of fat and muscle.
The results from the visual assessment are summarized in Table 3. Statistical significance was evaluated using a singlesample, twotailed ttest without applying any multiplecomparisonbased correction. From the total score in the table, it can be observed that the proposed method with D → ∞ performs on par with CD, whereas PPFM and D → ∞ are assigned somewhat lower scores and Noise2Void, BM3D, and WGANVGG exhibit the worst performances. The fact that D → ∞ performs on par with CD is remarkable because the latter is a supervised method, and therefore has access to more data during training. Furthermore, the proposed method has a smaller number of parameters than CD, which is reflected in its shorter inference time (Table 2). It can also be noted that the value D = 128 gives the optimal LPIPS, which measures the image fidelity in the feature space, whereas the visual assessment gives a higher score to the D → ∞ method because radiologists prefer its stronger noise suppression. The fact that adjusting D allows the optimization of the method for different performance metrics underscores the flexibility of the proposed method.
Discussion
Because image denoising was of interest, the simplest possible data consistency (or regularization) step was selected: the identity map. An interesting direction for future research is to determine how well the proposed method generalizes to other inverse problems. It can be surmised that updating only the regularization step is sufficient, as is the case with similar techniques based on diffusion models [34, 35]. One interesting application is the combination of image denoising and superresolution. However, it is possible that one also needs to update the hijacking method, because injecting the condition image directly into the sampling process may fail to generalize beyond the problem of image denoising. Even within the two tasks of denoising LDCT and PCCT images, noise characteristics vary widely. CT images are routinely reconstructed using different kernels, slice thicknesses, fields of view, and matrix sizes. These factors may have resulted in reconstructed images with significantly different noise characteristics. As shown in ref. [47], this may adversely affect the performance of image denoising techniques. Hence, it is possible that one needs to update the hyperparameters in the sampling algorithm, including the consistency step, to attain good performance over a wide range of settings.
Another interesting avenue for future research is the extension of the proposed method to 3D denoising. Given the structure of CT data, 2D denoising discards an abundance of rich information by not considering adjacent slices. It is possible that this additional information can aid in the recovery of more details observed in NDCT images from LDCT data. In particular, using information from adjacent slices may help better differentiate noise from signals. Extending the proposed method to 3D denoising has two effects. First, the network can be retrained using 3D data. The benefit of this approach is that it allows the network to optimize the use of information from adjacent slices. However, a disadvantage of this approach is that it requires retraining. Another possibility is to update only the data consistency step. One can keep the learned prior retained from the 2D data, thus not requiring any retraining, and combined with a regularization step that utilizes information from adjacent slices. Finally, the two approaches were combined. In future studies, the proposed method will be extended for 3D denoising.
In the current generation of PCCT scanners, spectral (materially resolved) CT are usually of interest. Hence, extending this method to spectral cases is of interest. Similar to the 3D case, one could allow the network to accept more channels, where the channels could be materialbasis images or virtual monoenergetic images at two or more energy levels. Another possibility is to train two separate networks for two different material bases or virtual monoenergetic images at different energy levels.
Although strong correlation between LPIPS [41] and visual image quality can be observed, it should be noted that LPIPS is computed using a network trained on RGB natural images. Relative to natural images, grayscale CT images should be considered outside the distribution. Nevertheless, the results indicate that using these selection criteria yields visually pleasing images in a LDCT dataset. Furthermore, the results for the PCCT data indicate that the network with these sampling hyperparameters generalized well. These empirical findings strongly suggest the intrinsic nature of LPIPS features. One possible way to make LPIPS more interpretable is to retrain the underlying network in LPIPS on a large dataset of CT/PCCT images. This important work is beyond the scope of this study and is left to future research.
Conclusions
In practice, paired data are typically unavailable for denoising CT images. In this study, an unsupervised version of the PPFM [19] was proposed and it was demonstrated that despite imposing a significantly laxer data requirement, there is only a small drop in the overall performance. To achieve this, a PFGM++ [29] trained in an unsupervised manner for unconditional generation with a sampling scheme that enforces consistency was combined with the input or condition image to enable sampling from the desired posterior. The proposed method includes a corresponding method based on diffusion models (EDM [27]), and it was demonstrated that PFGM++, with D as an additional hyperparameter, yields significant performance gains. The results indicate competitive performance compared with popular supervised, unsupervised, and nonDLbased image denoising techniques, including stateoftheart diffusionstyle models with NFE = 1, consistency models, clinical LDCT data, and clinical images from a prototype PCCT system.
Availability of data and materials
Code used for this project is available at: https://github.com/dennishein/pfgmpp_PCCT_denoising. Mayo lowdose CT data used are available at: https://aapm.app.box.com/s/eaw4jddb53keg1bptavvvd1sf4x3pe9h/folder/144226105715.
Abbreviations
 CT:

Computed tomography
 PCCT:

Photoncounting CT
 DL:

Deep learning
 NFE:

Number of function evaluations
 ODE:

Ordinary differential equation
 CD:

Consistency distillation
 PFGM:

Poisson flow generative models
 AAPM:

American Association of Physicists in Medicine
 NDCT:

Normaldose CT
 LDCT:

Lowdose CT
 LPIPS:

Learned Perceptual Image Patch Similarity
 ROI:

Regionofinterest
 SSIM:

Structural similarity index
 PSNR:

Peak signaltonoise ratio
 CNN:

Convolutional neural network
References
Wang G, Ye JC, De Man B (2020) Deep learning for tomographic image reconstruction. Nat Mach Intell 2(12):737–748. https://doi.org/10.1038/s4225602000273z
Koetzier LR, Mastrodicasa D, Szczykutowicz TP, van der Werf NR, Wang AS, Sandfort V et al (2023) Deep learning image reconstruction for CT: technical principles and clinical prospects. Radiology 306(3):e221257. https://doi.org/10.1148/radiol.221257
Willemink MJ, Persson M, Pourmorteza A, Pelc NJ, Fleischmann D (2018) Photoncounting CT: technical principles and clinical prospects. Radiology 289(2):293–312. https://doi.org/10.1148/radiol.2018172656
Flohr T, Petersilka M, Henning A, Ulzheimer S, Ferda J, Schmidt B (2020) Photoncounting CT review. Phys Med 79:126–136. https://doi.org/10.1016/j.ejmp.2020.10.030
Danielsson M, Persson M, Sjölin M (2021) Photoncounting xray detectors for CT. Phys Med Biol 66(3):03TR01. https://doi.org/10.1088/13616560/abc5a5
Higashigaito K, Euler A, Eberhard M, Flohr TG, Schmidt B, Alkadhi H (2022) Contrastenhanced abdominal CT with clinical photoncounting detector CT: assessment of image quality and comparison with energyintegrating detector CT. Acad Radiol 29(5):689–697. https://doi.org/10.1016/j.acra.2021.06.018
Chen H, Zhang Y, Kalra MK, Lin F, Chen Y, Liao PX et al (2017) Lowdose CT with a residual encoderdecoder convolutional neural network. IEEE Trans Med Imaging 36(12):2524–2535. https://doi.org/10.1109/TMI.2017.2715284
Wolterink JM, Leiner T, Viergever MA, Išgum I (2017) Generative adversarial networks for noise reduction in lowdose CT. IEEE Trans Med Imaging 36(12):2536–2545. https://doi.org/10.1109/TMI.2017.2708987
Yang QS, Yan PK, Zhang YB, Yu HY, Shi YY, Mou XQ et al (2018) Lowdose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss. IEEE Trans Med Imaging 37(6):1348–1357. https://doi.org/10.1109/TMI.2018.2827462
Kim B, Han M, Shim H, Baek J (2019) A performance comparison of convolutional neural networkbased image denoising methods: the effect of loss functions on lowdose CT images. Med Phys 46(9):3906–3923. https://doi.org/10.1002/mp.13713
Shan HM, Padole A, Homayounieh F, Kruger U, Khera RD, Nitiwarangkul C et al (2019) Competitive performance of a modularized deep neural network compared to commercial algorithms for lowdose CT image reconstruction. Nat Mach Intell 1(6):269–276. https://doi.org/10.1038/s4225601900579
Kim K, Soltanayev S, Chun SY (2020) Unsupervised training of denoisers for lowdose CT reconstruction without fulldose ground truth. IEEE J Sel Top Sign Process 14(6):1112–1125. https://doi.org/10.1109/JSTSP.2020.3007326
Yuan NM, Zhou J, Qi JY (2020) Half2half: deep neural network based CT image denoising without independent reference data. Phys Med Biol 65(21):215020. https://doi.org/10.1088/13616560/aba939
Li ZH, Zhou SW, Huang JZ, Yu LF, Jin MW (2021) Investigation of lowdose CT image denoising using unpaired deep learning methods. IEEE Trans Radiat Plasma Med Sci 5(2):224–234. https://doi.org/10.1109/TRPMS.2020.3007583
Wang S, Yang YR, Yin ZY, Wang AS (2023) Noise2Noise for denoising photon counting CT images: generating training data from existing scans. In: Proceedings of SPIE 12463, medical imaging 2023: physics of medical imaging, SPIE, San Diego, 7 April 2023. https://doi.org/10.1117/12.2653975
Niu C, Li MZ, Fan FL, Wu WW, Guo XD, Lyu Q et al (2023) Noise suppression with similaritybased selfsupervised deep learning. IEEE Trans Med Imaging 42(6):1590–1602. https://doi.org/10.1109/TMI.2022.3231428
Liu X, Xie YQ, Cheng J, Diao SH, Tan S, Liang XK (2023) Diffusion probabilistic priors for zeroshot lowdose CT image denoising. arXiv preprint arXiv: 2305.15887
Tivnan M, Teneggi J, Lee TC, Zhang RQ, Boedeker K, Cai L et al (2023) Notice of removal: Fourier diffusion models: a method to control MTF and NPS in scorebased stochastic image generation. IEEE Trans Med Imaging. https://doi.org/10.1109/TMI.2023.3335339
Hein D, Holmin S, Szczykutowicz T, Maltz JS, Danielsson M, Wang G et al (2023) PPFM: Image denoising in photoncounting CT using singlestep posterior sampling Poisson flow generative models. arXiv preprint arXiv: 2312.09754
Yu LF, Shiung M, Jondal D, McCollough CH (2012) Development and validation of a practical lowerdosesimulation tool for optimizing computed tomography scan protocols. J Comput Assisted Tomogr 36(4):477–487. https://doi.org/10.1097/RCT.0b013e318258e891
Huber NR, Missert AD, Gong H, Leng S, Yu LF, McCollough CH (2023) Technical note: phantombased training framework for convolutional neural network CT noise reduction. Med Phys 50(2):821–830. https://doi.org/10.1002/mp.16093
SohlDickstein J, Weiss EA, Maheswaranathan N, Ganguli S (2015) Deep unsupervised learning using nonequilibrium thermodynamics. In: Proceedings of the 32nd international conference on machine learning, JMLR.org, Lille, 611 July 2015
Ho J, Jain A, Abbeel P (2020) Denoising diffusion probabilistic models. In: Proceedings of the 34th international conference on neural information processing systems, Curran Associates Inc., Vancouver, 612 December 2020
Nichol AQ, Dhariwal P (2021) Improved denoising diffusion probabilistic models. In: Proceedings of the 38th international conference on machine learning, PMLR, Virtual, 1824 July 2021
Song Y, SohlDickstein J, Kingma DP, Kumar A, Ermon S, Poole B (2021) Scorebased generative modeling through stochastic differential equations. In: Proceedings of the 9th international conference on learning representations, OpenReview.net, Virtual, 37 May 2021
Song JM, Meng CL, Ermon S (2021) Denoising diffusion implicit models. In: Proceedings of the 9th international conference on learning representations, OpenReview.net, Virtual, 37 May 2021
Karras T, Aittala M, Aila T, Laine S (2022) Elucidating the design space of diffusionbased generative models. In: Proceedings of the 36th conference on neural information processing systems, NeurIPS, New Orleans, 28 November9 December 2022
Xu YL, Liu ZM, Tegmark M, Jaakkola T (2022) Poisson flow generative models. In: Proceedings of the 36th conference on neural information processing systems, NeurIPS, New Orleans, 28 November9 December 2022
Xu YL, Liu ZM, Tian YL, Tong SY, Tegmark M, Jaakkola T (2023) PFGM++: unlocking the potential of physicsinspired generative models. In: Proceedings of the 40th international conference on machine learning, PMLR, Honolulu, 2329 July 2023
Batzolis G, Stanczuk J, Schönlieb CB, Etmann C (2021) Conditional image generation with scorebased diffusion models. arXiv preprint arXiv: 2111.13606
Chung H, Sim B, Ye JC (2022) Comecloserdiffusefaster: accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of 2022 IEEE/CVF conference on computer vision and pattern recognition, IEEE, New Orleans, 1824 June 2022. https://doi.org/10.1109/CVPR52688.2022.01209
Saharia C, Chan W, Saxena S, Li LL, Whang J, Denton E et al (2022) Photorealistic texttoimage diffusion models with deep language understanding. In: Proceedings of the 36th conference on neural information processing systems, NeurIPS, New Orleans, 28 November9 December 2022
Saharia C, Chan W, Chang H, Lee C, Ho J, Salimans T et al (2022) Palette: Imagetoimage diffusion models. In: Proceedings of ACM SIGGRAPH 2022 conference proceedings, ACM, Vancouver, 24 July 2022. https://doi.org/10.1145/3528233.3530757
Yang S, Shen LY, Xing L, Ermon S (2022) Solving inverse problems in medical imaging with scorebased generative models. In: Proceedings of the tenth international conference on learning representations, OpenReview.net, Virtual, 2529 April 2022
Chung H, Lee ES, Ye JC (2023) MR image denoising and superresolution using regularized reverse diffusion. IEEE Trans Med Imaging 42(4):922–934. https://doi.org/10.1109/TMI.2022.3220681
Salimans T, Ho J (2022) Progressive distillation for fast sampling of diffusion models. In: Proceedings of the tenth international conference on learning representations, OpenReview.net, Virtual, 2529 April 2022
Song Y, Dhariwal P, Chen M, Sutskever I (2023) Consistency models. In: Proceedings of the 40th international conference on machine learning, PMLR, Honolulu, 2329 July 2023
Almqvist H, Crotty D, Nyren S, Yu J, ArnbergSandor F, Brismar T et al (2024) Initial clinical images from a secondgeneration prototype siliconbased photoncounting computed tomography system. Acad Radiol 31(2):572–581. https://doi.org/10.1016/j.acra.2023.06.031
AAPM (2016) Low dose CT grand challenge. https://www.aapm.org/grandchallenge/lowdosect/. Accessed 27 Mar 2024
Kingma DP, Ba J (2017) Adam: a method for stochastic optimization. arXiv preprint arXiv: 1412.6980
Zhang R, Isola P, Efros AA, Shechtman E, Wang O (2018) The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of 2018 IEEE/CVF conference on computer vision and pattern recognition, IEEE, Salt Lake City, 1823 June 2018. https://doi.org/10.1109/CVPR.2018.00068
Mäkinen Y, Azzari L, Foi A (2020) Collaborative filtering of correlated noise: exact transformdomain variance for improved shrinkage and patch matching. IEEE Trans Image Process 29:8339–8354. https://doi.org/10.1109/TIP.2020.3014721
Saharia C, Ho J, Chan W, Salimans T, Fleet DJ, Norouzi M (2023) Image superresolution via iterative refinement. IEEE Trans Pattern Anal Mach Intell 45(4):4713–4726
Krull A, Buchholz TO, Jug F (2019) Noise2Void  learning denoising from single noisy images. In: Proceedings of 2019 IEEE/CVF conference on computer vision and pattern recognition, IEEE, Long Beach, 1520 June 2019. https://doi.org/10.1109/CVPR.2019.00223
Wang Z, Bovik AC, Sheikh HR, Simoncelli EP (2004) Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process 13(4):600–612. https://doi.org/10.1109/TIP.2003.819861
Krizhevsky A (2014) One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv: 1404.5997
Huber NR, Missert AD, Yu LF, Leng S, McCollough CH (2021) Evaluating a convolutional neural network noise reduction method when applied to CT images reconstructed differently than training data. J Comput Assisted Tomogr 45(4):544–551. https://doi.org/10.1097/RCT.0000000000001150
Acknowledgements
The authors would like to thank Jimmy Yu and Torkel Brismar for taking part in the visual assessment of image quality.
Funding
This study was financially supported by MedTechLabs, GE HealthCare, the Swedish Research council, No. 2021–05103; and the Göran Gustafsson foundation, No. 2114.
Author information
Authors and Affiliations
Contributions
DH developed the proposed method and drafted the manuscript; SH oversaw the clinical study at Karolinska Insitutet and revised the manuscript; TS oversaw the clinical study at University of Wisconsin–Madison and revised the manuscript; GW contributed to discussion, experimental design and manuscript revision; MP conducted the reader study and revised the manuscript; JM and MD revised the manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
The clinical study was performed in accordance with the Declaration of Helsinki and approved by Swedish Ethics Review Agency 2020–04638 and 2021–01092 and prospectively consented IRB review UWIRB: 2022–1043.
Competing interests
DH discloses research collaboration with GE HealthCare.
MP discloses research support and license fees, GE HealthCare.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Hein, D., Holmin, S., Szczykutowicz, T. et al. Noise suppression in photoncounting computed tomography using unsupervised Poisson flow generative models. Vis. Comput. Ind. Biomed. Art 7, 24 (2024). https://doi.org/10.1186/s42492024001756
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s42492024001756