Distilling Photon-Counting CT into Routine Chest CT through Clinically Validated Degradation Modeling
Authors: Junqi Liu, Xinze Zhou, Wenxuan Li, Scott Ye, Arkadiusz Sitek, Xiaofeng Yang, Yucheng Tang, Daguang Xu, Kai Ding, Kang Wang, Yang Yang, Alan L. Yuille, Zongwei Zhou
Organizations: Johns Hopkins University · University of California, San Francisco · Harvard Medical School · Emory University · Nvidia · Johns Hopkins Medicine
Photon-counting CT (PCCT) provides superior image quality with higher spatial resolution and lower noise compared to conventional energy-integrating CT (EICT), but its limited clinical availability restricts large-scale research and clinical deployment. To bridge this gap, we propose SUMI, a simulated degradation-to-enhancement method that learns to reverse realistic acquisition artifacts in low-quality EICT by leveraging high-quality PCCT as reference. Our central insight is to explicitly model realistic acquisition degradations, transforming PCCT into clinically plausible lower-quality counterparts and learning to invert this process. The simulated degradations were validated for clinical realism by board-certified radiologists, enabling faithful supervision without requiring paired acquisitions at scale. As outcomes of this technical contribution, we: (1) train a latent diffusion model on 1,046 PCCTs, using an autoencoder first pre-trained on both these PCCTs and 405,379 EICTs from 145 hospitals to extract general CT latent features that we release for reuse in other generative medical imaging tasks; (2) construct a large-scale dataset of over 17,316 publicly available EICTs enhanced to PCCT-like quality, with radiologist-validated voxel-wise annotations of airway trees, arteries, veins, lungs, and lobes; and (3) demonstrate substantial improvements: across external data, SUMI outperforms state-of-the-art image translation methods by 15% in SSIM and 20% in PSNR, improves radiologist-rated clinical utility in reader studies, and enhances downstream top-ranking lesion detection performance, increasing sensitivity by up to 15% and F1 score by up to 10%. Our results suggest that emerging imaging advances can be systematically distilled into routine EICT using limited high-quality scans as reference.
Figures & tables
Figure 1: Overview of SUMI. (1.) A continual learning autoencoder is pre-trained on over 400,000 CT volumes from 145 hospitals across 19 countries under pixel-wise loss Lpp and adversarial loss Ldis , making it the largest and most geographically diverse open-source medical CT autoencoder to date, overpassing prior methods (LDM [ 27 ] : 0, MedDiff [ 35 ] : 2, MONAI [ 5 ] : 7, Med3D [ 7 ] : 8, MAISI [ 39 ] : 10). (2.) A clinical verified degradation simulator transforms high-quality PCCT into three realistic low-quality counterparts: sparse-view (reducing 2D projections), low-dose (reducing photon counts), and conventional (reducing resolution while increasing noise and artifacts), covering the primary sources of EICT degradation. (3.) SUMI takes a degraded input x~i from step (2.) and passes it through the autoencoder from step (1.) to produce an enhanced output xi′ . Training uses the original PCCT xi as ground truth under four losses: pixel-wise loss Lpp for structural fidelity, segmentation loss Lseg and HU consistency loss LHU to preserve organ boundaries and tissue densities, such as pancreas, tumor, vessel, and airway tree using pre-computed masks SEG(⋅) , and adversarial loss Ldis via discriminator D(⋅) for image realism. Once trained, SUMI can enhance any EICT scan to PCCT-like quality without retraining.
method
sparse view
low dose
conventional
mixed
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SR3 [ 28 ]
46.2
8.7
42.0
8.8
44.4
8.6
47.0
8.7
Pix2Pix [ 17 ]
43.6
11.9
60.0
22.6
60.8
18.1
62.9
22.7
Swin2SR [ 9 ]
62.1
26.7
59.9
28.0
29.0
24.9
66.8
26.4
NLM [ 4 ]
73.2
28.4
72.3
30.2
40.9
26.1
76.9
27.0
SRGAN [ 19 ]
56.8
23.2
85.5
33.0
82.8
32.4
63.1
26.5
Table 1: SUMI improves image quality across degradations. On 446 external patients, our method surpasses the second-best method by up to +35.9% SSIM and +19.5% PSNR under sparse-view, low-dose, conventional, and mixed settings. Baselines are ordered by ascending average performance.
Figure 2: SUMI preserves anatomical structure and tissue density. Pearson correlation ( r ) between ground truth and enhanced CT measurements across all organs shows that SUMI maintains anatomical accuracy and HU consistency. Organ masks are obtained with VISTA3D [ 16 ] .
dataset
DSB17 [ 18 ]
LIDC-IDRI [ 1 ]
LNDb19 [ 25 ]
LTRC [ 2 ]
scans ( N )
1,596
1,018
324
1,496
dataset
Luna16 [ 29 ]
MIDRC [ 36 ]
NLST [ 32 ]
RSNA-STR [ 8 ]
scans ( N )
854
10,496
422
1,110
Table 2: Our method processes and enhances a massive, multi-source cohort of 17,316 public chest CT scans. Cohorts are listed in alphabetical order.
Figure 3: SUMI demonstrates superior generalization and enhancement quality. For each example, the left image shows the CT slice and the right shows the airway tree segmentation. (a) Cross-dataset evaluation: compared with autoencoders trained with limited medical data scale (LDM, Med3D, MAISI), SUMI better preserves fine airway topology under domain shift. (b) PCCT enhancement with ground truth: compared with strong baselines (NLM [ 4 ] , SRGAN [ 19 ] , NEED [ 11 ] ), SUMI maintains structural integrity and small airway branches more faithfully.
Luna16 [ 29 ] (N=222)
LNDb19 [ 25 ] (N=224)
DSB17 [ 18 ] (N=198)
lung nodule, MONAI 1 [ 5 ]
lung nodule, MONAI 1 [ 5 ]
lung tumor, grt123 2 [ 20 ]
method
Sen.
Spec.
F1
AUC
Sen.
Spec.
F1
AUC
Sen.
Spec.
F1
AUC
baseline
75.8
88.4
81.6
88.0
61.5
86.9
55.2
74.0
78.4
82.1
80.2
87.0
+ SUMI
84.5
88.4
87.2
92.5
75.0
89.2
70.4
84.5
88.2
86.4
87.3
91.8
Δ
+8.7
+0.0
+5.6
+4.5
+13.5
+2.3
+15.2
+10.5
+9.8
+4.3
+7.1
+4.8
Table 3: SUMI improves downstream detection across three independent chest CT cohorts. When integrated into open-source SOTA baselines, our enhancement consistently increases performance on Luna16 [ 29 ] , LNDb19 [ 25 ] , and DSB17 [ 18 ] , achieving up to +15.2% F1 and +10.5% AUC. Baselines are trained on standard-quality CT, while “+ SUMI” denotes the same architectures retrained using our PCCT-like enhanced data.
Figure 4: SUMI improves chest CT quality across datasets. Quality score distributions (by a pretrained scorer, 0–1, higher is better) on Luna16 [ 29 ] , LNDb19 [ 25 ] , DSB17 [ 18 ] , and a private PCCT dataset. On public datasets, enhanced scans consistently shift toward higher scores, indicating robust quality gains. On PCCT, a controlled degradation-to-enhancement benchmark shows that SUMI narrows the gap to real high-quality PCCT standards.
method
sparse view
low dose
conventional
mixed
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
w/o. sparse view
71.9
30.2
90.0
35.2
87.2
35.3
73.2
30.1
w/o. low dose
72.7
30.2
82.6
28.0
84.3
32.9
76.2
31.5
w/o. conventional
75.8
27.0
86.0
33.0
85.3
32.0
81.9
32.4
all degrades
91.6
33.0
92.8
37.3
88.2
35.0
90.2
33.7
Table 4: Ablation of degradation simulations in SUMI. Leave-one-out experiments under identical settings show that removing sparse-view or low-dose simulation degrades performance on an external test set (446 scans), confirming their role in robust generalization.
Computed tomography (CT) images are frequently degraded by acquisition artifacts, including noise, blur, streaking, aliasing, and metal artifacts. Yet CT enhancement is still largely evaluated using image quality metrics with limited perceptual and clinical validity, while existing datasets remain focused on isolated restoration tasks, hindering unified benchmarking across diverse degradation types. We present CT-DegradBench, a dataset and benchmark for CT degradation detection and severity estimation under controlled single- and mixed-artifact settings. CT-DegradBench enables systematic evaluation across multiple degradation families and severity levels within a common experimental framework. We further propose SeSpeCT (Semantic-Spectral CT degradation estimation), a framework that combines semantic priors from medical vision-language models with complementary frequency-domain cues for artifact analysis. SeSpeCT constructs a training-free semantic quality axis in the multimodal embedding space using radiology-informed text prompts, without task-specific fine-tuning, and combines it with spectral features that capture degradation-specific frequency patterns. The resulting representation enables joint prediction of artifact type and severity. Experimental results show that SeSpeCT consistently outperforms the evaluated baselines under both single- and mixed-degradation settings. The framework is available at https://github.com/yousranb/CT-DEGRADBENCH.
Yousra Nabila Taifour, Marouane Tliba, Zuheng Ming +9
1Université Sorbonne Paris Nord · University of Dubai · 3Northwestern University +3
Cone-beam computed tomography (CBCT) images are problematic in clinical medicine because of their low contrast and high artifact content compared with conventional CT images. Although there are some studies to improve image quality, in regions subject to organ deformation, the anatomical structure may change after such image quality improvement. In this study, we propose an overcorrection-free CBCT image quality enhancement method based on a conditional latent diffusion model using pseudo-CBCT images. Pseudo-CBCT images are created from CT images using a simple method that simulates CBCT artifacts and are spatially consistent with the CT images. By performing self-supervised learning with these spatially consistent paired images, we can improve image quality while maintaining anatomical structures. Furthermore, extending the framework of the conditional diffusion model to latent space improves the efficiency of image processing. Our model was trained on pelvic CT-pseudo-CBCT paired data and was applied to both pseudo-CBCT and real CBCT data. The experimental results using data of 75 cases show that with our proposed method, the structural changes were less than 1/1000th (in terms of the number of pixels) of those of a conventional method involving learning with real images, and the correlation coefficient between the CT value distributions of the generated and reference images was 0.916, approaching the same level as conventional methods. We also confirmed that the proposed framework achieves faster processing and superior improvement performance compared with the framework of a conditional diffusion model, even under constrained training settings.
Reconstructing 3D Computed Tomography (CT) images from a few X-ray projections is a highly ill-posed inverse problem due to the loss of volumetric information. We propose PhyDiCT, a training-free framework that integrates a differentiable Physics-based forward model, grounded in the Beer-Lambert law, with a text-conditioned Diffusion as a strong prior to reconstruct 3D lung CT images. We refer to our approach as training-free since the prior model is used without fine-tuning, and our goal is to steer the denoising procedure to generate samples consistent with X-ray observations. We guide the diffusion generation using Split Gibbs sampling to jointly optimize for projection fidelity (reward) and consistency with prior knowledge. Also, we introduce a test-time refinement step that enhances image realism and anatomical coherence. We extensively evaluate our method on publicly available 3D CT datasets using both perceptual and semantic metrics, demonstrating that it surpasses existing plug-and-play diffusion and fully trained reconstruction approaches. Our findings highlight that combining a strong generative prior with the underlying physics of image formation substantially improves reconstruction quality, e.g., 7.5% improvement on SSIM compared to full training methods. Code will be released at https://github.com/batmanlab/PhyDiCT.