Metal implants cause severe artifacts that degrade CT image quality and diagnostic accuracy. Conventional deep learning-based metal artifact reduction (MAR) methods typically adopt an end-to-end mapping in either the image or projection domain. However, these approaches often exhibit instability and inaccuracy because attenuation images depend on scanning protocols and equipment specifications. Moreover, existing MAR methods do not fully exploit the additional spectral information from spectral CT to enhance artifact reduction in conventional single-energy CT. To address these limitations, we propose a Spectral Prior Enhanced Consistency Network for MAR (s-MAR). Unlike conventional methods that learn energy-dependent attenuation images, s-MAR learns energy-invariant basis material images, integrating spectral priors and enforcing projection consistency to improve reconstruction fidelity. Specifically, s-MAR performs artifact reduction in the Basis Material Image Domain (BMID) by combining the Projection Restoration Network (PR-Net) and Basis Material Image Restoration Network (BR-Net) to produce artifact-suppressed energy-invariant water and bone basis images. To promote data fidelity, the water and bone basis images are forward-projected and input into a Spectral Projection Synthesis Network (SPS-Net), which estimates the multi-energy projections and calculates a projection consistency loss with the original projections. Finally, s-MAR employs the Image Reconstruction Network (IR-Net) to fuse the spectral prior enhanced basis material images and generate the target attenuation image. Since s-MAR performs MAR in the BMID, training on real dual-energy data enables s-MAR to implicitly learn spectral priors that guide MAR. Extensive simulation and clinical experiments demonstrate that s-MAR surpasses existing MAR methods in both visual quality and quantitative performance metrics.

MRI has become essential in clinical diagnosis due to its high resolution and multiple contrast mechanisms. However, the relatively long acquisition time limits its broader application. To address this issue, this study presents an innovative conditional guided diffusion model, named TC-KANRecon, which incorporates the Multi-Free U-KAN module and a dynamic clipping strategy. TC-KANRecon model aims to accelerate the MRI reconstruction process through deep learning methods while maintaining the reconstruction quality. The MF-UKAN module can effectively balance the tradeoff between image denoising and structure preservation. Specifically, it presents the multi-head attention mechanisms and scalar modulation factors, which significantly enhance the model's robustness and structure preservation capabilities in complex noise environments. Moreover, the dynamic clipping strategy in TC-KANRecon adjusts the cropping interval according to the sampling steps, thereby mitigating image detail loss while preserving the visual features of the images. Furthermore, the Conditional Guidance Model incorporates full-sampling k-space information, realizing efficient fusion of conditional information, enhancing the model's ability to process complex data, and improving the realism and detail richness of reconstructed images. Experimental results demonstrate that the proposed method outperforms other MRI reconstruction methods in both qualitative and quantitative evaluations. Notably, TC-KANRecon method exhibits excellent reconstruction results when processing high-noise, low-sampling-rate MRI data.

Sparse-view computed tomography (SCT) has great potential for enabling routine clinical low-dose imaging. However, SCT reconstruction is an ill-posed inverse problem, in which images reconstructed by FBP algorithm suffer from severe streaking artifacts and noise. To solve these problems, deep-learning-based sinogram interpolation methods have been proposed to estimate missing projections, but the interpolation errors are often amplified during reconstruction, resulting in secondary artifacts. In this study, we propose a pure view-by-view back-projection tensor (PVBP-Tensor) interpolation framework with the rotation-equivariant convolution for SCT reconstruction, named PVBP-IREConv. The PVBP-Tensor serves as an intermediate representation that retains angular sampling properties and explicit structural features. Meanwhile, utilizing the PVBP-Tensor inherent similar radial-like structures, PVBP-IREConv first adopts rotation-equivariant convolution to efficiently synthesize the missing projections. The reconstructed image is then obtained by summing all slices of the synthetic full-view PVBP-Tensor along the angular dimension, thereby localizing the interpolation errors to corresponding image pixels and preventing secondary artifacts. Moreover, since the PVBP-IREConv operates directly on PVBP-Tensor obtained after back-projection, it naturally generalizes across diverse scanning geometries and reconstruction parameters within a unified network. Experimental results on both fan-beam and helical sparse-view CT demonstrate that PVBP-IREConv achieves superior image quality and quantitative performance compared with existing methods.

OBJECTIVES:To address the limitations of data storage and transfer caused by exponential growth of medical imaging data size, we propose a frequency-adaptive implicit neural compression (FAINC) method for medical images based on optimized implicit neural networks (INRs). METHODS:A retrospective analysis was conducted on abdominal CT data from 356 patients in the KiTS19 and AVT datasets. We developed the FAINC method, a multi-subnetwork collaborative compression framework, which first evaluates the frequency-domain complexity of image blocks using the Spectral Sparsity Index (SSI), and then dynamically allocates them to subnetworks of different capacities through a frequency-domain gating mechanism. Compression output is achieved by combining parameter quantization with entropy coding. To assess its performance, the proposed method was compared with mainstream commercial compression standards (H.265/HEVC and JPEG2000), the implicit neural representation method NeRV, and the deep-learning-based compression method DVC. RESULTS:The FAINC method achieved the best reconstruction performance on both KiTS19 and AVT datasets at high compression ratios. At bitrates of BPV=0.32 and BPV=0.34, the FAINC method obtained the highest PSNR (47.03 and 50.76), the highest SSIM (0.9853 and 0.9930), and the lowest RMSE (0.0045 and 0.0029), achieving also a significantly higher subjective image quality score than other methods. Ablation studies demonstrated that the frequency-domain gating mechanism and dynamic parameter allocation contributed approximately 2.05 dB and 1.87 dB PSNR improvements, respectively. CONCLUSIONS:The proposed method substantially enhances image reconstruction quality at high compression ratios and outperforms the existing mainstream approaches in terms of structural fidelity and compression efficiency. The FAINC method provides a promising technical solution for efficient storage and low-bandwidth remote transfer of medical image data.

Recent advances in deep-learning-based methods have shown great potential in improving low-dose CT image quality. Meanwhile, these methods are constructed based on a large, centralized, and diverse CT dataset from multiple institutions that is difficult to collect and share due to the high-cost acquisition and data privacy regulations. Previously developed federated learning (FL)-based methods enable collaborative and decentralized training without exchanging local data to preserve data privacy. In this work, we focus on analyzing the robustness of FL-based methods against dataset shifts (i.e., the datasets among multiple institutions are from different scanners, different protocols, or different sampling conditions). The results show that the FL-based CT reconstruction methods are sensitive to domain shifts, which can be attributed to the data heterogeneity among multiple institutions. Based on these findings, we propose a unified CT reconstruction method that leverages high-quality metadata (e.g., low-dose images and their corresponding normal-dose counterparts) stored on the cloud server to address the challenge of multi-institutional domain shifts. For simplicity, we refer to the proposed method as FM-iRadonMAP, representing federated metadata learning (FMDL) with a personalized condition-modulated iRadonMAP (CM-iRadonMAP). Specifically, the FM-iRadonMAP consists of two modules, i.e., CM-iRadonMAP and FMDL. CM-iRadonMAP introduces the knowledge of client-specific sampling conditions, i.e., imaging geometries and scan protocols, into iRadonMAP reconstruction network at each client to modulate the reconstruction effectively. FMDL trains a supervised meta model using high-quality metadata in an additional round and then adaptively unifies the network parameters of the meta model with those of the local models from all clients for broadcasting, addressing the issue of data heterogeneity. A large-scale multi-institutional CT dataset is used to validate and evaluate the reconstruction performance of the FM-iRadonMAP. The experimental results demonstrate the feasibility of the FM-iRadonMAP for multi-institutional CT reconstruction with severe data heterogeneity.

The clinical translation of AI in medical imaging faces critical challenges including scare annotated data, long-tailed pathological distributions, and privacy constraints in virtual imaging trials. To address these limitations, we propose a Multi-conditional Diffusion framework with Texture Constraints (MDTC) for synthesizing clinically reliable lesions in CT images. The key innovation lies in the joint integration of anatomical mask guidance and a Gray-Level Co-occurence Matrix-based texture classifier into the diffusion network. This dual-constraint mechanism uniquely enforces structural fidelity and pathologically heterogeneous characteristics in synthetic lesions, effectively preserves pathological heterogeneity and simultaneously enhances the authenticity of the generated lesionst in existing methods. The experiments on hepatocellular carcinoma and pulmonary nodule datasets demonstrate that the proposed MDTC method achieves favorable performance in terms of texture fidelity, a significant improvement in PSNR, and notable optimization in FID compared to classic generation models. Furthermore, downstream classifiers trained on synthetic data remain capable of maintaining the vast majority of baseline classification performance, while data augmentation for rare pulmonary nodules significantly improved the classification efficacy. This work establishes a new paradigm for generating diagnostically meaningful synthetic data, effectively alleviating critical bottlenecks in virtual imaging trial. The experimental results demonstrate that MDTC-generated lesions preserve structural authenticity and have potential to mitigate texture homogenization, thereby enabling robust downstream diagnostic model training and offering a practical, privacy preserving solution to expand virtual imaging trials toward underrepresented disease groups and real world clinical workflows.

Background:Identifying predictive markers for immunotherapy in non-small cell lung cancer (NSCLC) is critical for personalized treatment. This study aimed to construct a predictive model that integrates clinical features, enhanced computed tomography (CT)-radiomics, and deep learning (DL) features for the assessment of durable clinical benefit (DCB) from immunotherapy in patients with advanced NSCLC and to provide biological interpretability to predictions by integrating radiogenomic data. Methods:We conducted a retrospective analysis of 201 advanced NSCLC patients who underwent immunotherapy with CT images, with data supplemented from The Cancer Imaging Archive (TCIA). Radiomics features (RFs) were extracted from enhanced CT images, and DL features were derived using a pre-trained ResNet-34 model. DCB-related signatures were constructed using the least absolute shrinkage and selection operator (LASSO) algorithm, and fusion nomogram models were developed by integrating significant clinical variables, radiomics, and DL features. Shapley additive explanations were employed to quantify the impact of radiomics-DL features on model predictions. Gene set enrichment and biological correlation analyses based on transcriptomic TCIA data were performed to explore the biological significance of radiomics-DL score. Results:Statistically significant clinical predictors included initial efficacy, brain metastases, programmed death-ligand 1 (PD-L1) expression, and hemoglobin levels. The fusion nomogram model demonstrated the highest predictive accuracy for DCB, with area under the curve (AUC) values of 0.843 in the train cohort and 0.894 in the test cohort, surpassing individual feature sets. Biological exploration revealed associations between radiomics-DL score and biological characteristics, including immune responses and immunoregulation. Conclusions:This integrated approach shows the potential of combining clinical, radiomics and deep learning features (DLFs) as a noninvasive biomarker for predicting immunotherapy efficacy in NSCLC, assisting in patient selection and clinical decision-making. Radiotranscriptomic analysis may reveal key cellular and immune patterns associated with radiomics-DL signature.

In computed tomography imaging, metal implants frequently generate severe artifacts that compromise image quality and hinder diagnostic accuracy. There are three main challenges in the existing methods: the deterioration of organ and tissue structures, dependence on sinogram data, and an imbalance between resource use and restoration efficiency. Addressing these issues, we introduce MARMamba, which effectively eliminates artifacts caused by metals of different sizes while maintaining the integrity of the original anatomical structures of the image. Furthermore, this model only focuses on CT images affected by metal artifacts, thus negating the requirement for additional input data. The model is a streamlined UNet architecture, which incorporates multi-scale Mamba (MS-Mamba) as its core module. Within MS-Mamba, a flip mamba block captures comprehensive contextual information by analyzing images from multiple orientations. Subsequently, the average maximum feed-forward network integrates critical features with average features to suppress the artifacts. This combination allows MARMamba to reduce artifacts efficiently. The experimental results demonstrate that our model excels in reducing metal artifacts, offering distinct advantages over other models. It also strikes an optimal balance between computational demands, memory usage, and the number of parameters, highlighting its practical utility in the real world. The code of the presented model is available at: https://github.com/RICKand-MORTY/MARMamba.

In this study, we systematically evaluated the iodine quantification accuracy and image noise suppression capabilities of a deep learning reconstruction algorithm (ClearInfinity, CI) under 60 kVp ultra-low tube voltage computed tomography (CT) conditions, comparing it with filtered backprojection (FBP) and hybrid iterative reconstruction (ClearView, CV). A CT performance phantom containing inserts with varying iodine concentrations (40, 28, 22, 12, 6, 3, and 2 mg/mL) was scanned six times (60 kVp, 386 mA) using a NeuViz Epoch Elite CT scanner. Images were reconstructed using FBP, CV (at 20%, 40%, 60%, and 80% intensities), and CI (at equal intensity). CT values, image noise (standard deviation SD), and coefficients of variation (cv) of the iodine inserts were measured. Absolute percentage bias (APB) and contrast-to-noise ratio (CNR) were calculated. Results show that CI achieved optimal quantitative accuracy at 40% reconstruction intensity and provided the strongest noise reduction at 80%, with a maximum SD reduction of up to 79.59%. At all intensity levels, CI significantly outperformed CV and FBP in terms of APB, noise suppression (especially at low iodine concentrations), measurement stability, and CNR. These findings confirm that CI is an effective solution for producing low-noise, low-bias, and highly stable images in ultra-low-dose CT.

Metal implants introduce severe artifacts in CT images, compromising diagnostic reliability. Supervised metal artifact reduction (MAR) models trained on simulated data are effective but often fail due to domain gaps when applied to real clinical data. Unsupervised methods trained on real images avoid such gaps but suffer from weak artifact suppression and training instability. To address these challenges, we propose d-MAR, a novel MAR framework that performs diffusion-driven domain transformations between simulated and real image domains. Specifically, real image domain (RID) data is transformed into the simulated image domain (SID), processed by a MAR model trained on simulation-paired data, and transformed back into RID. We harness diffusion models as a transformation bridge and introduce two targeted conditional sampling techniques-conditional input and sampling enhancement-based on Fourier-extracted low-frequency image components. This enables domain alignment without random generation, ensuring consistent anatomical fidelity. The proposed d-MAR can reduce real metal artifacts originating from different scanning protocols and devices with a MAR model trained with simulated paired data. Evaluations on Clinical Head, Clinical Body, and dental CBCT datasets show that d-MAR consistently outperforms conventional MAR methods in both quantitative metrics and visual quality, demonstrating strong generalization capability.

Purpose.Physical computed tomography (CT) perfusion (CTP) phantoms are useful for controlled protocol evaluation, but a modular three-dimensional (3D)-printable platform that can generate distinguishable compartment-specific dynamic contrast behavior and support repeatability assessment remains needed. We developed the separate module uniting phantom (SMUPhantom) for protocol sensitivity studies and workflow benchmarking.Methods.The SMUPhantom consists of three modules representing inflow, exchange, and outflow regions. Polyoxymethylene beads, soft wax, and acrylonitrile-butadiene-styrene inserts were used to create healthy, penumbra-mimicking, and infarct-core-mimicking pathways, and a detachable bypass-flow pathway was added for flow-routing experiments. Two experimental series were performed on a NeuViz Epoch Elite CT scanner: contrast volume and injection rate were varied separately, each with and without bypass flow. Time-density curves and workstation-derived cerebral blood flow, cerebral blood volume (CBV), mean transit time (MTT), and time to peak (TTP) were evaluated. Five repeat measurements per condition assessed intra-scanner repeatability.Results.The three compartments showed separable dynamic responses, with the healthy pathway exhibiting earlier and higher enhancement and the infarct-core-mimicking pathway showing lower and delayed responses. Increasing contrast volume broadened the bolus response and raised peak enhancement from 220.8 to 454.2 HU in the healthy pathway and from 149.4 to 354.4 HU in the infarct-core-mimicking pathway, whereas increasing injection rate shortened TTP from 22-25 s to 12-15 s and MTT from 18-21 s to 10-14 s. Opening bypass flow modified the compartment curves and map appearance, particularly in the infarct-core-mimicking pathway, while preserving interpretable trends across repeated scans. Repeatability analysis showed stable peak enhancement and CBV measurements, with CVs not exceeding 3.54% and 2.81%, respectively.Conclusion.SMUPhantom provides a 3D-printable and quantitatively characterized platform for CTP protocol sensitivity studies, repeatability assessment, and workflow benchmarking under controlled flow-routing conditions.

Linear models are extensively employed in tomographic imaging to model the imaging system, with tomographic reconstruction from measurement data typically formulated as a linear inverse problem. However, linear models fail to accurately characterize the imaging system under nonideal imaging conditions. Nonlinear models provide a superior characterization of tomographic imaging systems in such conditions, albeit posing substantial challenges in directly solving the associated nonlinear inverse problem. This article introduces a physics-inspired dual regression network (Tomo-Net), which integrates dual learning with the physical mechanisms of tomographic imaging to solve the nonlinear inverse problem. Tomo-Net constructs a dual regression network by incorporating both primal and dual mappings. Primal mapping facilitates the transformation of degraded measurement data into high-quality reconstructed images, while dual mapping enables the transformation of reconstructed images back to the original measurement data. Tomo-Net incorporates the physical mechanisms of tomographic imaging in both mappings to constrain the solution of the nonlinear inverse problem, and solves the issue of nonuniqueness in the solution space by enforcing data fidelity constraints, dual cycle consistency constraints, and physical mechanism consistency constraints. In addition, Tomo-Net applies consistency constraints during the training phase rather than inference, thereby avoiding additional inference costs. While adaptable to various imaging modalities with suitable modifications, this article focuses on X-ray computed tomography (CT) imaging as a specific example and evaluates its effectiveness across various reconstruction tasks. The extensive experimental results demonstrate that Tomo-Net outperforms traditional machine learning models in solving nonlinear inverse problems in CT. The Tomo-Net code is publicly available at https://github.com/guoyii/Tomo-Net

Model-based image reconstruction methods with regularization terms have been developed to suppress noise in the photon-counting detector CT (PCD-CT) images. Meanwhile, most regularization terms are usually designed based on the image characteristics, but do not account for the real-world noise distribution within the images, which may produce undesired biases in the reconstruction results. In this work, we analyze the noise characteristics of PCD-CT images, i.e., material dependent, spatial location dependent, and energy dependent characteristics, and present a Noise Characterized Model-based Iterative Reconstruction (NCM-IR) method for high-fidelity PCD-CT imaging. Specifically, a three-dimensional total variation (3DTV) is utilized to describe the texture characteristics in the PCD-CT images. Then, the characteristics of noise are modeled in an explicit form with a universal approximator, i.e., Gaussian mixture model (GMM). Moreover, both the GMM and 3DTV are introduced into the presented NCM-IR method. Finally, in the presented NCM-IR, we optimize the parameters of the noise distribution and 3DTV with respect to the reconstruction accuracy by using a designed Expectation Maximization algorithm. The presented NCM-IR method is extensively evaluated in numerical and preclinical studies and we demonstrate the presented NCM-IR method with characterized noise distribution outperforms the competing methods that either utilize only characterized noise distribution or lack it entirely in terms of noise reduction, structure preservation, and material decomposition accuracy.

With the increasing use of computed tomography (CT), concerns about radiation dose have grown. Deep-learning-based methods have shown great promise in improving low-dose CT image quality while further reducing patient dose. However, most deep-learning-based methods are trained on vendor-specific CT datasets with varying imaging conditions and dose levels, which results in poor generalizability across vendors due to marked data heterogeneity. Moreover, the centralization of multicenter datasets is restricted by the high costs of data collection and privacy regulations. To overcome these challenges, we propose FedM2CT, a federated metadata-constrained method with mutual learning for all-in-one CT reconstruction. This method enables simultaneous reconstruction of multivendor CT images with different imaging geometries and sampling protocols in one framework. Specifically, FedM2CT consists of 3 modules: task-specific iRadonMAP (TS-iRadonMAP), condition-prompted mutual learning (CPML), and federated metadata learning (FMDL). TS-iRadonMAP performs task-specific low-dose reconstruction, CPML shares condition-prompted knowledge between clients and the server, and FMDL aggregates model parameters with a metamodel to effectively mitigate the effect of data heterogeneity. Extensive experiments under 3 different settings demonstrate that the proposed FedM2CT achieves outstanding results compared to other methods, both qualitatively and quantitatively, showing the potential to achieve the goal of all-in-one CT reconstruction with different low-dose tasks, i.e., low-milliampere-second, sparse-view, and limited-angle.

BACKGROUND:Neoadjuvant chemoimmunotherapy (NACI) has become one of the most widely adopted therapies for head and neck squamous cell carcinoma (HNSCC) before surgery. However, the accurate prediction of patients responding to this therapy has been challenging owing to the lack of predictive biomarkers. METHODS:In the present study, histologically confirmed HNSCC patients with complete clinicopathological data, who received chemotherapy plus programmed cell death protein 1 inhibitor as the NACI regimen for 2-3 cycles before radical surgery between 2021 and 2023, were screened, and both clinicopathological and magnetic resonance imaging (MRI) data were collected and divided into training, testing, and external validation cohorts. Both traditional radiomics and deep-learning techniques were employed to extract features from MRI, followed by feature selection using both Spearman correlation and least absolute shrinkage and selection operator analysis. The selected features were incorporated into predictive models using a logistic regression classifier for pathologic complete response. RESULTS:The results demonstrated that three out of seven features extracted from MRI were deep-learning features. Notably, the integration of deep-learning features with clinicopathological and radiomics features increased the area under the curve in the training, testing, and external validation cohorts to 0.781, 0.759, and 0.740, respectively. Moreover, multimodal prediction for patient stratification can significantly improve the prognosis of HNSCC patients undergoing NACI. CONCLUSIONS:In conclusion, deep-learning features from MRI can augment traditional imaging analysis to uncover hidden predictive patterns reflecting responsiveness to NACI in HNSCC patients, and integration of different data types provides a more robust prediction strategy.

The generalization of deep learning-based low-dose computed tomography (CT) reconstruction models to doses unseen in the training data is important and remains challenging. Previous efforts heavily rely on paired data to improve the generalization performance and robustness through collecting either diverse CT data for re-training or a few test data for fine-tuning. Recently, diffusion models have shown promising and generalizable performance in low-dose CT (LDCT) reconstruction, however, they may produce unrealistic structures due to the CT image noise deviating from Gaussian distribution and imprecise prior information from the guidance of noisy LDCT images. In this paper, we propose a noise-inspired diffusion model for generalizable LDCT reconstruction, termed NEED, which tailors diffusion models for noise characteristics of each domain. First, we propose a novel shifted Poisson diffusion model to denoise projection data, which aligns the diffusion process with the noise model in pre-log LDCT projections. Second, we devise a doubly guided diffusion model to refine reconstructed images, which leverages LDCT images and initial reconstructions to more accurately locate prior information and enhance reconstruction fidelity. By cascading these two diffusion models for dual-domain reconstruction, our NEED requires only normal-dose data for training and can be effectively extended to various unseen dose levels during testing via a time step matching strategy. Extensive qualitative, quantitative, and segmentation-based evaluations on two datasets demonstrate that our NEED consistently outperforms state-of-the-art methods in reconstruction and generalization performance. Source code is made available at https://github.com/qgao21/NEED.

BACKGROUND:Photon counting CT has demonstrated exceptional performance in spatial resolution, density resolution, and image quality, earning recognition as a groundbreaking technology in medical imaging. However, its technical implementation continues to face substantial challenges, including charge sharing effects. OBJECTIVE:To develop a spatio-energetic charge-sharing modulation model for accurate photon counting CT simulation (SmuSim). Specifically, SmuSim is built upon the previously developed photon counting toolkit (PcTK) and thoroughly incorporates the charge sharing effects that occur in photon counting CT. METHODS:The proposed SmuSim firstly enrolls three primary modules, i.e., photon transport, charge transport, and charge induction to characterize the charge sharing effects in the photon counting CT imaging chain. Then, Monte Carlo simulation is also conducted to validate the feasibility of the proposed SmuSim with well-built charge sharing effects model. RESULTS:Under diverse detector configurations, SmuSim's energy spectrum response curves exhibit a remarkable alignment with Monte Carlo simulations, in stark contrast to the Pctk results. In both digital and clinical phantom studies, SmuSim effectively simulates distorted photon counting CT images. In digital physical phantom simulations, the deviations in attenuation coefficient due to charge sharing effects are -49.70%, -19.66%, and -3.33% for the three energy bins, respectively. In digital clinical phantom simulations, the differences in attenuation coefficient are -19.92%, -4.98%, and -0.6%, respectively. In the two simulation studies, the deviations between the results obtained from SmuSim and those from Monte Carlo simulation are less than 3% and 2%, respectively, demonstrating the effectiveness of the proposed SmuSim. CONCLUSION:We analyze charge sharing effects in photon counting CT, a comprehensive analytical model, and finally simulate CT images with charge sharing effects for evaluation.

OBJECTIVES:We propose a low-dose CT image restoration method based on central guidance and alternating optimization (FedGP). METHODS:The FedGP framework revolutionizes the traditional federated learning model by adopting a structure without a fixed central server, where each institution alternatively serves as the central server. This method uses an institution-modulated CT image restoration network as the core of client-side local training. Through a federated learning approach of central guidance and alternating optimization, the central server leverages local labeled data to guide client-side network training to enhance the generalization capability of the CT imaging model across multiple institutions. RESULTS:In the low-dose and sparse-view CT image restoration tasks, the FedGP method showed significant advantages in both visual and quantitative evaluation and achieved the highest PSNR (40.25 and 38.84), the highest SSIM (0.95 and 0.92), and the lowest RMSE (2.39 and 2.56). Ablation study of FedGP demonstrated that compared with FedGP(w/o GP) without central guidance, the FedGP method better adapted to data heterogeneity across institutions, thus ensuring robustness and generalization capability of the model in different imaging conditions. CONCLUSIONS:FedGP provides a more flexible FL framework to solve the problem of CT imaging heterogeneity and well adapts to multi-institutional data characteristics to improve generalization ability of the model under diverse imaging geometric configurations.

Computed tomography perfusion (CTP) plays a crucial role in guiding reperfusion therapy and patient selection for acute ischemic stroke (AIS) through perfusion parameter maps of the brain; however, its widespread use is limited by the complexity of acquisition protocols and high radiation dose. Previous studies have attempted to reduce radiation exposure by equally lowering the temporal sampling rate; however, it may miss the peak of arterial enhancement, leading to underestimation of blood flow parameter. Here, we investigate the feasibility of using a generative adversarial network (GAN) to generate perfusion maps from 3 phases of CTP (mCTP). The three phases were chosen based on the multiphase computed tomography angiography scanning protocol: the peak arterial input function phase, the peak venous output function phase, and the delayed venous output function phase. The findings demonstrate that the GAN model achieved high visual overlap and performance for cerebral blood flow and time-to-maximum maps, with a mean structural similarity index measure of 0.921 to 0.971 and 0.817 to 0.883, a mean normalized root mean squared error of 0.019 to 0.108 and 0.058 to 0.064, and a mean learned perceptual image patch similarity of 0.039 to 0.088 and 0.141 to 0.146, respectively. For the 2 external datasets, the volume agreement between the model- and CTP-derived infarct and hypoperfusion areas was the intraclass correlation coefficient of 0.731 to 0.883 and 0.499 to 0.635, and the Spearman correlation coefficient of 0.720 to 0.808 and 0.533 to 0.6540, respectively. Qualitative assessments of diagnostic quality further confirmed that the mCTP-derived maps were comparable to those obtained from traditional CTP. In conclusion, the GAN-based model is effective in generating perfusion maps from mCTP, which could serve as a viable alternative to traditional CTP in the diagnostic evaluation of AIS.