Pub Date : 2025-10-20DOI: 10.1109/TCI.2025.3622903
Tao Hong;Zhaoyi Xu;Jason Hu;Jeffrey A. Fessler
Model-based iterative reconstruction plays a key role in solving inverse problems. However, the associated minimization problems are generally large-scale, nonsmooth, and sometimes even nonconvex, which present challenges in designing efficient iterative solvers. Preconditioning methods can significantly accelerate the convergence of iterative methods. In some applications, computing preconditioners on-the-fly is beneficial. Moreover, forward models in image reconstruction are typically represented as operators, and the corresponding explicit matrices are often unavailable, which brings additional challenges in designing preconditioners. Therefore, for practical use, computing and applying preconditioners should be computationally inexpensive. This paper adapts the randomized Nyström approximation to compute effective preconditioners that accelerate image reconstruction without requiring an explicit matrix for the forward model. We leverage modern GPU computational platforms to compute the preconditioner on-the-fly. Moreover, we propose efficient approaches for applying the preconditioners to problems with classical nonsmooth regularizers, i.e., wavelet, total variation, and Hessian Schatten-norm. Our numerical results on image deblurring, super-resolution with impulsive noise, and 2D computed tomography reconstruction illustrate the efficiency and effectiveness of the proposed preconditioner.
{"title":"Using Randomized Nyström Preconditioners to Accelerate Variational Image Reconstruction","authors":"Tao Hong;Zhaoyi Xu;Jason Hu;Jeffrey A. Fessler","doi":"10.1109/TCI.2025.3622903","DOIUrl":"https://doi.org/10.1109/TCI.2025.3622903","url":null,"abstract":"Model-based iterative reconstruction plays a key role in solving inverse problems. However, the associated minimization problems are generally large-scale, nonsmooth, and sometimes even nonconvex, which present challenges in designing efficient iterative solvers. Preconditioning methods can significantly accelerate the convergence of iterative methods. In some applications, computing preconditioners on-the-fly is beneficial. Moreover, forward models in image reconstruction are typically represented as operators, and the corresponding explicit matrices are often unavailable, which brings additional challenges in designing preconditioners. Therefore, for practical use, computing and applying preconditioners should be computationally inexpensive. This paper adapts the randomized Nyström approximation to compute effective preconditioners that accelerate image reconstruction without requiring an explicit matrix for the forward model. We leverage modern GPU computational platforms to compute the preconditioner on-the-fly. Moreover, we propose efficient approaches for applying the preconditioners to problems with classical nonsmooth regularizers, i.e., wavelet, total variation, and Hessian Schatten-norm. Our numerical results on image deblurring, super-resolution with impulsive noise, and 2D computed tomography reconstruction illustrate the efficiency and effectiveness of the proposed preconditioner.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1630-1643"},"PeriodicalIF":4.8,"publicationDate":"2025-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145560746","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-17DOI: 10.1109/TCI.2025.3623041
Sida Gao;Ziyang Li;Zhengyu Wu;Qi Li;Pengcheng Jia;Yutong Li;Shutian Liu;Zhengjun Liu
Single-shot digital holography has demonstrated great potential for capturing the complex amplitude of optical fields in lensless imaging systems. Nevertheless, accurately reconstructing intricate interference patterns remains challenging due to high-frequency texture distortion and limited generalization across diverse real-world conditions. To address these issues, we propose a physics-embedded network that explicitly incorporates angular spectrum propagation and spatial-domain amplitude constraints into a deep learning framework, ensuring both physical consistency and strong representational capacity. Building on this foundation, we devised HoloSSL, a physics-aware heterogeneous self-supervised learning framework for robust single-frame holographic reconstruction. The framework consists of two complementary stages: (1) physics-aware self-supervised pretraining using Hermite-Gaussian synthesized holograms to learn a prior-consistent mapping, and (2) unsupervised domain adaptation using real captured intensity-only holograms, where intensity fidelity and propagation consistency jointly guide label-free fine-tuning. To enhance texture representation, we design a swin-hourglass block that integrates cross-scale self-attention with frequency-aware modeling. Extensive simulations and real-world experiments demonstrate that HoloSSL outperforms state-of-the-art methods in terms of reconstruction fidelity, structural consistency, and robustness to noise, providing a new paradigm for adaptive, interpretable, and high-fidelity holographic imaging.
{"title":"High-Fidelity Holographic Reconstruction via Physics-Aware Heterogeneous Self-Supervised Learning","authors":"Sida Gao;Ziyang Li;Zhengyu Wu;Qi Li;Pengcheng Jia;Yutong Li;Shutian Liu;Zhengjun Liu","doi":"10.1109/TCI.2025.3623041","DOIUrl":"https://doi.org/10.1109/TCI.2025.3623041","url":null,"abstract":"Single-shot digital holography has demonstrated great potential for capturing the complex amplitude of optical fields in lensless imaging systems. Nevertheless, accurately reconstructing intricate interference patterns remains challenging due to high-frequency texture distortion and limited generalization across diverse real-world conditions. To address these issues, we propose a physics-embedded network that explicitly incorporates angular spectrum propagation and spatial-domain amplitude constraints into a deep learning framework, ensuring both physical consistency and strong representational capacity. Building on this foundation, we devised HoloSSL, a physics-aware heterogeneous self-supervised learning framework for robust single-frame holographic reconstruction. The framework consists of two complementary stages: (1) physics-aware self-supervised pretraining using Hermite-Gaussian synthesized holograms to learn a prior-consistent mapping, and (2) unsupervised domain adaptation using real captured intensity-only holograms, where intensity fidelity and propagation consistency jointly guide label-free fine-tuning. To enhance texture representation, we design a swin-hourglass block that integrates cross-scale self-attention with frequency-aware modeling. Extensive simulations and real-world experiments demonstrate that HoloSSL outperforms state-of-the-art methods in terms of reconstruction fidelity, structural consistency, and robustness to noise, providing a new paradigm for adaptive, interpretable, and high-fidelity holographic imaging.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1469-1478"},"PeriodicalIF":4.8,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145455859","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-17DOI: 10.1109/TCI.2025.3623006
Yichen Zhou;Yong Wang
The acquisition of high-resolution images of space orbiting satellites by spaceborne ISAR has significant importance for enhancing space situation awareness. However, challenges persist with the increase of resolution. Two main issues arise: (1) The assumption of a stable imaging projection plane is invalid under the current imaging regime. (2) The high-order spatial-variant range cell migration and phase errors caused by complex relative motion is introduced. Therefore, it is important to optimize the existing ISAR imaging geometry models and imaging algorithms. In this paper, integration of high-order motion compensation and high-resolution imaging algorithm is proposed. Firstly, the spaceborne ISAR geometric imaging model is established, and the rationality of the model and the unique properties application to image space satellites target are investigated. Subsequently, by effectively utilizing the properties, a high-order range migration elimination method based on image partition is proposed, along with the image partition parameters estimation method. For the remaining phase error, a high-resolution imaging algorithm is introduced to eliminate error, which can overcome the influence of cross terms and strong scattering points. Finally, compared with the existing algorithms, simulation results validate the effectiveness and superiority of the proposed algorithm under different SNR.
{"title":"Joint High-Order Motion Compensation and Imaging Algorithm for Ultrahigh-Resolution Spaceborne ISAR","authors":"Yichen Zhou;Yong Wang","doi":"10.1109/TCI.2025.3623006","DOIUrl":"https://doi.org/10.1109/TCI.2025.3623006","url":null,"abstract":"The acquisition of high-resolution images of space orbiting satellites by spaceborne ISAR has significant importance for enhancing space situation awareness. However, challenges persist with the increase of resolution. Two main issues arise: (1) The assumption of a stable imaging projection plane is invalid under the current imaging regime. (2) The high-order spatial-variant range cell migration and phase errors caused by complex relative motion is introduced. Therefore, it is important to optimize the existing ISAR imaging geometry models and imaging algorithms. In this paper, integration of high-order motion compensation and high-resolution imaging algorithm is proposed. Firstly, the spaceborne ISAR geometric imaging model is established, and the rationality of the model and the unique properties application to image space satellites target are investigated. Subsequently, by effectively utilizing the properties, a high-order range migration elimination method based on image partition is proposed, along with the image partition parameters estimation method. For the remaining phase error, a high-resolution imaging algorithm is introduced to eliminate error, which can overcome the influence of cross terms and strong scattering points. Finally, compared with the existing algorithms, simulation results validate the effectiveness and superiority of the proposed algorithm under different SNR.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1479-1493"},"PeriodicalIF":4.8,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145455915","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Single-photon Lidar imaging offers a significant advantage in 3D imaging due to its high resolution and long-range capabilities, however it is challenging to apply in noisy environments with multiple targets per pixel. To tackle these challenges, several methods have been proposed. Statistical methods demonstrate interpretability on the inferred parameters, but they are often limited in their ability to handle complex scenes. Deep learning-based methods have shown superior performance in terms of accuracy and robustness, but they lack interpretability or they are limited to a single-peak per pixel. In this paper, we propose a deep unrolling algorithm for dual-peak single-photon Lidar imaging. We introduce a hierarchical Bayesian model for multiple targets and propose a neural network that unrolls the underlying statistical method. To support multiple targets, we adopt a dual depth maps representation and exploit geometric deep learning to extract features from the point cloud. The proposed method takes advantages of statistical methods and learning-based methods in terms of accuracy and quantifying uncertainty. The experimental results on synthetic and real data demonstrate the competitive performance when compared to existing methods, while also providing uncertainty information.
{"title":"Graph Attention-Driven Bayesian Deep Unrolling for Dual-Peak Single-Photon Lidar Imaging","authors":"Kyungmin Choi;JaKeoung Koo;Stephen McLaughlin;Abderrahim Halimi","doi":"10.1109/TCI.2025.3623000","DOIUrl":"https://doi.org/10.1109/TCI.2025.3623000","url":null,"abstract":"Single-photon Lidar imaging offers a significant advantage in 3D imaging due to its high resolution and long-range capabilities, however it is challenging to apply in noisy environments with multiple targets per pixel. To tackle these challenges, several methods have been proposed. Statistical methods demonstrate interpretability on the inferred parameters, but they are often limited in their ability to handle complex scenes. Deep learning-based methods have shown superior performance in terms of accuracy and robustness, but they lack interpretability or they are limited to a single-peak per pixel. In this paper, we propose a deep unrolling algorithm for dual-peak single-photon Lidar imaging. We introduce a hierarchical Bayesian model for multiple targets and propose a neural network that unrolls the underlying statistical method. To support multiple targets, we adopt a dual depth maps representation and exploit geometric deep learning to extract features from the point cloud. The proposed method takes advantages of statistical methods and learning-based methods in terms of accuracy and quantifying uncertainty. The experimental results on synthetic and real data demonstrate the competitive performance when compared to existing methods, while also providing uncertainty information.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1494-1504"},"PeriodicalIF":4.8,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145455808","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Hyperspectral imaging empowers machine vision systems with the distinct capability of identifying materials through recording their spectral signatures. Recent efforts in data-driven spectral reconstruction aim at extracting spectral information from RGB images captured by cost-effective RGB cameras, instead of dedicated hardware. Published work reports exceedingly high numerical scores for this reconstruction task, yet real-world performance lags substantially behind. In this paper we systematically analyze the performance of such methods with three groups of dedicated experiments. First, we evaluate the practical overfitting limitations with respect to current datasets by training the networks with less data, validating the trained models with unseen yet slightly modified data, and cross-dataset validation. Second, we reveal fundamental limitations in the ability of RGB to spectral methods to deal with metameric or near-metameric conditions, which have so far gone largely unnoticed due to the insufficiencies of existing datasets. We achieve this by validating the trained models with metamer data generated by metameric black theory and re-training the networks with various forms of metamers. This methodology can also be used for data augmentation as a partial mitigation of the dataset issues, although the RGB to spectral inverse problem remains fundamentally ill-posed. Finally, we analyze the potential for modifying the problem setting to achieve better performance by exploiting some form of optical encoding provided by either incidental optical aberrations or some form of deliberate optical design. Our experiments show that such approaches do indeed provide improved results under certain circumstances, however their overall performance is limited by the same dataset issues as in the plain RGB to spectral scenario. We therefore conclude that future progress on snapshot spectral imaging will heavily depend on the generation of improved datasets which can then be used to design effective optical encoding strategies.
{"title":"Limitations of Data-Driven Spectral Reconstruction: An Optics-Aware Analysis","authors":"Qiang Fu;Matheus Souza;Eunsue Choi;Suhyun Shin;Seung-Hwan Baek;Wolfgang Heidrich","doi":"10.1109/TCI.2025.3622928","DOIUrl":"https://doi.org/10.1109/TCI.2025.3622928","url":null,"abstract":"Hyperspectral imaging empowers machine vision systems with the distinct capability of identifying materials through recording their spectral signatures. Recent efforts in data-driven spectral reconstruction aim at extracting spectral information from RGB images captured by cost-effective RGB cameras, instead of dedicated hardware. Published work reports exceedingly high numerical scores for this reconstruction task, yet real-world performance lags substantially behind. In this paper we <italic>systematically analyze</i> the performance of such methods with three groups of dedicated experiments. First, we evaluate the practical overfitting limitations with respect to current datasets by training the networks with less data, validating the trained models with unseen yet slightly modified data, and cross-dataset validation. Second, we reveal <italic>fundamental limitations</i> in the ability of RGB to spectral methods to deal with metameric or near-metameric conditions, which have so far gone largely unnoticed due to the insufficiencies of existing datasets. We achieve this by validating the trained models with metamer data generated by metameric black theory and re-training the networks with various forms of metamers. This methodology can also be used for data augmentation as a partial mitigation of the dataset issues, <italic>although the RGB to spectral inverse problem remains fundamentally ill-posed</i>. Finally, we analyze the potential for modifying the problem setting to achieve better performance by exploiting some form of optical encoding provided by either incidental optical aberrations or some form of deliberate optical design. Our experiments show that such approaches do indeed provide improved results under certain circumstances, however their overall performance is limited by the same dataset issues as in the plain RGB to spectral scenario. We therefore conclude that <italic>future progress on snapshot spectral imaging will heavily depend on the generation of improved datasets which can then be used to design effective optical encoding strategies</i>.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1505-1520"},"PeriodicalIF":4.8,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11206531","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145455956","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Diffractive lens (DLs) computational imaging promises lightweight infrared imaging. However, it is still challenging to accurately obtain the degradation of DLs computational imaging systems to achieve high-quality imaging. Degradation estimation methods based on image-to-image are the most feasible way, but existing methods face the problem of low accuracy for severe degradation estimation. In this paper, we proposed a severe degradation estimation framework based on KernelNet, and realized the improvement of the accuracy of severe degenerate estimation by more than 1.13 dB. Furthermore, DyUnet, an image restoration network, was proposed, which is based on dynamic convolution. The results of the mid-wave diffractive lens infrared computational imaging system indicated that our method achieved high-quality imaging results comparable to the conventional one, while the number of optical elements was reduced from 7 to 2, and the weight was reduced by 50$%$.
{"title":"High-Quality Diffractive Lens Infrared Computational Imaging via Severe Degradation Estimation Framework","authors":"Pengzhou Ji;Xiong Dun;Yujie Xing;Xuquan Wang;Jian Zhang;Hongmei Li;Kan Zhao;Hongfei Jiao;Zhanshan Wang;Xinbin Cheng","doi":"10.1109/TCI.2025.3619811","DOIUrl":"https://doi.org/10.1109/TCI.2025.3619811","url":null,"abstract":"Diffractive lens (DLs) computational imaging promises lightweight infrared imaging. However, it is still challenging to accurately obtain the degradation of DLs computational imaging systems to achieve high-quality imaging. Degradation estimation methods based on image-to-image are the most feasible way, but existing methods face the problem of low accuracy for severe degradation estimation. In this paper, we proposed a severe degradation estimation framework based on KernelNet, and realized the improvement of the accuracy of severe degenerate estimation by more than 1.13 dB. Furthermore, DyUnet, an image restoration network, was proposed, which is based on dynamic convolution. The results of the mid-wave diffractive lens infrared computational imaging system indicated that our method achieved high-quality imaging results comparable to the conventional one, while the number of optical elements was reduced from 7 to 2, and the weight was reduced by 50<inline-formula><tex-math>$%$</tex-math></inline-formula>.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1458-1468"},"PeriodicalIF":4.8,"publicationDate":"2025-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11197645","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145405368","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-09DOI: 10.1109/TCI.2025.3619821
Shanzhou Niu;Shuo Li;Tinghua Wang;Weiwen Wu;You Zhang;Jing Wang;Jianhua Ma
Cerebral perfusion computed tomography (CPCT) can non-invasively and rapidly assess blood flow circulation in the brain, making it widely adopted in clinical settings. However, the dynamic scanning protocol associated with CPCT entails substantial ionizing radiation exposure, leading to elevated radiation risks. Lowering the tube current can efficiently reduce radiation dose, but leads to significant image quality deterioration for standard filtered back-projection (FBP) algorithm due to increased quantum noise in measured projection data. In this study, we present an iterative image reconstruction method to improve the low-dose CPCT image quality, which uses the prior image constrained total variation-stokes (PICTVS) based on the penalized weighted least squares (PWLS) criterion. This method leverages information from the prior image to enhance the image quality of low-dose CPCT. Specifically, PICTVS utilizes high-quality geometric structural information from the prior image and fuses it into low-dose CPCT image reconstruction while preserving the main features of the target image. An effective alternating minimization method was developed to solve the objective function associated with the PWLS-PICTVS reconstruction. The novelty of the PWLS-PICTVS algorithm is listed as follows: (1) The PICTVS regularization incorporates the structural information of prior image into the target image where the gradients of both images align; (2) In image areas where the gradients differ, the PICTVS regularization employs total variation (TV) instead; and (3) The PICTVS regularization facilitates the integration of shared edge structure information from a high-quality prior image into the low-dose image while avoiding introducing mismatched anatomy information. Qualitative and quantitative analyses were conducted to assess the efficacy of the PWLS-PICTVS image reconstruction algorithm using a digital brain perfusion phantom and simulated low-dose clinical patient data. The experimental results show that the PWLS-PICTVS algorithm significantly improves noise suppression, streak artifact reduction, and edge preservation when compared with the other reconstruction methods. Importantly, the CPCT images reconstructed using the PWLS-PICTVS method yield more accurate hemodynamic parameter maps, enhancing their potential for clinical diagnosis.
{"title":"Prior Image Constrained Total Variation-Stokes for Cerebral Perfusion CT Imaging","authors":"Shanzhou Niu;Shuo Li;Tinghua Wang;Weiwen Wu;You Zhang;Jing Wang;Jianhua Ma","doi":"10.1109/TCI.2025.3619821","DOIUrl":"https://doi.org/10.1109/TCI.2025.3619821","url":null,"abstract":"Cerebral perfusion computed tomography (CPCT) can non-invasively and rapidly assess blood flow circulation in the brain, making it widely adopted in clinical settings. However, the dynamic scanning protocol associated with CPCT entails substantial ionizing radiation exposure, leading to elevated radiation risks. Lowering the tube current can efficiently reduce radiation dose, but leads to significant image quality deterioration for standard filtered back-projection (FBP) algorithm due to increased quantum noise in measured projection data. In this study, we present an iterative image reconstruction method to improve the low-dose CPCT image quality, which uses the prior image constrained total variation-stokes (PICTVS) based on the penalized weighted least squares (PWLS) criterion. This method leverages information from the prior image to enhance the image quality of low-dose CPCT. Specifically, PICTVS utilizes high-quality geometric structural information from the prior image and fuses it into low-dose CPCT image reconstruction while preserving the main features of the target image. An effective alternating minimization method was developed to solve the objective function associated with the PWLS-PICTVS reconstruction. The novelty of the PWLS-PICTVS algorithm is listed as follows: (1) The PICTVS regularization incorporates the structural information of prior image into the target image where the gradients of both images align; (2) In image areas where the gradients differ, the PICTVS regularization employs total variation (TV) instead; and (3) The PICTVS regularization facilitates the integration of shared edge structure information from a high-quality prior image into the low-dose image while avoiding introducing mismatched anatomy information. Qualitative and quantitative analyses were conducted to assess the efficacy of the PWLS-PICTVS image reconstruction algorithm using a digital brain perfusion phantom and simulated low-dose clinical patient data. The experimental results show that the PWLS-PICTVS algorithm significantly improves noise suppression, streak artifact reduction, and edge preservation when compared with the other reconstruction methods. Importantly, the CPCT images reconstructed using the PWLS-PICTVS method yield more accurate hemodynamic parameter maps, enhancing their potential for clinical diagnosis.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1447-1457"},"PeriodicalIF":4.8,"publicationDate":"2025-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145351909","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Sparse-view CT reconstruction represents a prototypical ill-posed inverse problem. The implementation of deep learning solutions has proven to be highly successful in this field. The dual-domain reconstruction network achieves a favorable trade-off between reconstruction performance and computational cost by leveraging the powerful mapping capability of deep learning and the domain-transform relying on analytical reconstruction algorithms. However, further research is required to enhance the domain-transform in this field. Inspired by the successful utilization of low-rank prior in various medical imaging tasks, we proposed an end-to-end one-shot dual-domain network for sparse-view CT reconstruction. The domain-transform was designed as a high-fidelity multi-channel parallel back-projection in proposed network. In this way, feature maps between channels in the image domain imply strong low-rank priors. We implemented the singular value thresholding algorithm as a network layer, learning parameters and thresholds in a data-driven manner, fully leveraging the low-rank prior across channels to greatly reduce information loss and distortion during domain-transform. Moreover, we constructed a projection completion network based on dual attention mechanism that synthesizes missing view projections by effectively utilizing potential local correlation among projection domains during fan-beam scanning. In the image domain, a refine subnetwork based on Vision Transformer utilizes mix-scale features to implement two-dimensional filtering belonging to the back-projection filter algorithm. Extensive experiments on two clinically relevant datasets have demonstrated that the proposed network achieves competing performance on both quantitative metrics and visual quality.
{"title":"DROLL: Dual-Domain Reconstruction Network With a High-Fidelity Domain-Transform Operator Based on Learned Low-Rank Prior for Sparse-View CT Reconstruction","authors":"Haowen Zhang;Pengcheng Zhang;Yikun Zhang;Yang Chen;Yi Liu;Zhiguo Gui","doi":"10.1109/TCI.2025.3617255","DOIUrl":"https://doi.org/10.1109/TCI.2025.3617255","url":null,"abstract":"Sparse-view CT reconstruction represents a prototypical ill-posed inverse problem. The implementation of deep learning solutions has proven to be highly successful in this field. The dual-domain reconstruction network achieves a favorable trade-off between reconstruction performance and computational cost by leveraging the powerful mapping capability of deep learning and the domain-transform relying on analytical reconstruction algorithms. However, further research is required to enhance the domain-transform in this field. Inspired by the successful utilization of low-rank prior in various medical imaging tasks, we proposed an end-to-end one-shot dual-domain network for sparse-view CT reconstruction. The domain-transform was designed as a high-fidelity multi-channel parallel back-projection in proposed network. In this way, feature maps between channels in the image domain imply strong low-rank priors. We implemented the singular value thresholding algorithm as a network layer, learning parameters and thresholds in a data-driven manner, fully leveraging the low-rank prior across channels to greatly reduce information loss and distortion during domain-transform. Moreover, we constructed a projection completion network based on dual attention mechanism that synthesizes missing view projections by effectively utilizing potential local correlation among projection domains during fan-beam scanning. In the image domain, a refine subnetwork based on Vision Transformer utilizes mix-scale features to implement two-dimensional filtering belonging to the back-projection filter algorithm. Extensive experiments on two clinically relevant datasets have demonstrated that the proposed network achieves competing performance on both quantitative metrics and visual quality.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1375-1390"},"PeriodicalIF":4.8,"publicationDate":"2025-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145315292","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-02DOI: 10.1109/TCI.2025.3617237
Qi Wang;Yufang Cai;Haijun Yu;Fenlin Liu;Weiwen Wu
Limited data computed tomography (LDCT) plays a critical role in accelerating the scanning process and reducing radiation exposure for patients. However, LDCT reconstruction is inherently an ill-posed inverse problem, often resulting in pronounced edge artifacts and the loss of fine structural details. In recent years, score-based generative models (SGMs) have shown great promise in LDCT reconstruction by alleviating the ill-posedness and enabling high-fidelity image recovery in the case of noise-free condition. However, in practical CT systems, measurement data is often contaminated by noise. The coexistence of noise and limited data presents significant challenges for SGM-based image reconstruction methods. To address this challenge, this study proposes a Model-Informed Stable Diffusion (MISD) model which integrates a sampling process with a generative prior in the image-space module and a physics prior in the projection-space module. The projection-space module incorporates physical information to establish a noise suppression mechanism, effectively reducing the impact of noise. At the same time, the image-space module uses a generative model to progressively reconstruct clear structures and features from an initial state characterized by pure noise. Together, these two modules form a cohesive mathematical framework, utilizing iterative optimization to gradually minimize the effects of noise and artifacts. Experimental results show that the MISD method consistently achieves higher quantitative metrics and recovers finer structural details than other state-of-the-art reconstruction techniques, both on simulated and real datasets.
{"title":"MISD: Model-Informed Stable Diffusion Model for Limited Noisy Data CT Reconstruction","authors":"Qi Wang;Yufang Cai;Haijun Yu;Fenlin Liu;Weiwen Wu","doi":"10.1109/TCI.2025.3617237","DOIUrl":"https://doi.org/10.1109/TCI.2025.3617237","url":null,"abstract":"Limited data computed tomography (LDCT) plays a critical role in accelerating the scanning process and reducing radiation exposure for patients. However, LDCT reconstruction is inherently an ill-posed inverse problem, often resulting in pronounced edge artifacts and the loss of fine structural details. In recent years, score-based generative models (SGMs) have shown great promise in LDCT reconstruction by alleviating the ill-posedness and enabling high-fidelity image recovery in the case of noise-free condition. However, in practical CT systems, measurement data is often contaminated by noise. The coexistence of noise and limited data presents significant challenges for SGM-based image reconstruction methods. To address this challenge, this study proposes a Model-Informed Stable Diffusion (MISD) model which integrates a sampling process with a generative prior in the image-space module and a physics prior in the projection-space module. The projection-space module incorporates physical information to establish a noise suppression mechanism, effectively reducing the impact of noise. At the same time, the image-space module uses a generative model to progressively reconstruct clear structures and features from an initial state characterized by pure noise. Together, these two modules form a cohesive mathematical framework, utilizing iterative optimization to gradually minimize the effects of noise and artifacts. Experimental results show that the MISD method consistently achieves higher quantitative metrics and recovers finer structural details than other state-of-the-art reconstruction techniques, both on simulated and real datasets.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1434-1446"},"PeriodicalIF":4.8,"publicationDate":"2025-10-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145352017","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Coded Aperture Snapshot Spectral Imagers (CASSI) are designed to acquire 2D coded acquisitions of hyperspectral scenes (HSS). Associated with reconstruction algorithms, they allow to analyze the HSS using a small amount of data. Various configurations of such imagers exist, with different distortions. In this article we show that reconstruction quality is relatively insensitive to these distortions, if known and correctly modeled. To this end, we introduce a differentiable ray-tracing-based model that incorporates aberrations and distortions to render coded hyperspectral acquisitions using CASSI. Such simulated acquisitions are used to train state-of-the-art hyperspectral cube reconstruction algorithms. We also adapted these algorithms through the use of a ray-tracing-based mapping function which accounts for aberrations and distortions. We evaluated four comparable CASSI systems with varying degree of optical aberrations and misalignments, using five state-of-the-art hyperspectral cube reconstruction algorithms. Our analyses show that if known and properly modeled, the effects of geometric distortions of the system and misalignment of the dispersive elements have a marginal impact on the overall reconstruction quality. Therefore, relaxing traditional constraints on measurement conformity and fidelity to the scene enables the development of novel imaging instruments, guided by performance metrics applied to the design or the processing of acquisitions. By providing a complete framework for design, simulation and evaluation, this work contributes to the optimization and exploration of new CASSI systems and more generally to the computational imaging community.
{"title":"The Marginal Importance of Known Distortions and Alignment in CASSI Systems","authors":"Léo Paillet;Antoine Rouxel;Hervé Carfantan;Simon Lacroix;Antoine Monmayrant","doi":"10.1109/TCI.2025.3617235","DOIUrl":"https://doi.org/10.1109/TCI.2025.3617235","url":null,"abstract":"Coded Aperture Snapshot Spectral Imagers (CASSI) are designed to acquire 2D coded acquisitions of hyperspectral scenes (HSS). Associated with reconstruction algorithms, they allow to analyze the HSS using a small amount of data. Various configurations of such imagers exist, with different distortions. In this article we show that reconstruction quality is relatively insensitive to these distortions, if known and correctly modeled. To this end, we introduce a differentiable ray-tracing-based model that incorporates aberrations and distortions to render coded hyperspectral acquisitions using CASSI. Such simulated acquisitions are used to train state-of-the-art hyperspectral cube reconstruction algorithms. We also adapted these algorithms through the use of a ray-tracing-based mapping function which accounts for aberrations and distortions. We evaluated four comparable CASSI systems with varying degree of optical aberrations and misalignments, using five state-of-the-art hyperspectral cube reconstruction algorithms. Our analyses show that if known and properly modeled, the effects of geometric distortions of the system and misalignment of the dispersive elements have a marginal impact on the overall reconstruction quality. Therefore, relaxing traditional constraints on measurement conformity and fidelity to the scene enables the development of novel imaging instruments, guided by performance metrics applied to the design or the processing of acquisitions. By providing a complete framework for design, simulation and evaluation, this work contributes to the optimization and exploration of new CASSI systems and more generally to the computational imaging community.","PeriodicalId":56022,"journal":{"name":"IEEE Transactions on Computational Imaging","volume":"11 ","pages":"1391-1403"},"PeriodicalIF":4.8,"publicationDate":"2025-10-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145315293","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：