Yang Chen

Objective. Real-time x-ray fluoroscopy (XRF) is widely used for navigating coronary interventions. X-ray coronary angiography (XCA) is needed to facilitate the interventional procedure, resulting in excessive consumption of iodinated contrast agents. To reduce the contrast consumption, the vessel masks derived from XCA images are overlaid on XRF images, assuming periodic heartbeats. To address the two-dimensional (2D) nature of XCA limits, using 3D/2D coronary artery registration, centerlines with depth information derived from three-dimensional (3D) coronary computed tomography angiography (CCTA) can provide the necessary depth guidance and discriminate vessel overlapping. However, the existing tree-aligning feature-based registration strategy suffers from the error accumulation introduced by cascading multi-step processes or time-consuming limitation of iterative optimization in the 3D/2D tree alignment step. Approach. In this work, we propose an intensity-based end-to-end 3D/2D tree alignment method to address these limitations by removing the image-to-graph modality transformation in the intraoperative pipeline. Specifically, we design AlignNet for the intensity-based 3D/2D tree alignment, RegisNet for the centerline registration with graph regularization. AlignNet and RegisNet are trained using a cost-efficient supervised learning strategy based on large-scale simulated coronary deformations. Three cascaded end-to-end networks including nnUnet, AlignNet, and RegisNet are integrated into the intraoperative pipeline. Main results. Numerical simulations involving qualitative and quantitative analyses of large-scale CCTA–XCA samples achieved a mean registration error of 0.42 mm. Clinical validation using paired CCTA–XCA data from six patients achieved a mean registration error of 0.359 mm and a mean processing speed of 16.3 fps. In addition, validation using one paired CCTA–XCA–XRF case demonstrated the feasibility of generating depth-guided coronary roadmaps for interventional navigation. Significance. These results demonstrate that the proposed technique has the potential to achieve accurate and real-time depth-guided coronary intervention navigation, and it is valuable for subsequent multi-center clinical studies.

Non-ideal measurement computed tomography (CT) employs suboptimal imaging protocols to expand CT applications. However, the resulting trade-offs degrade image quality, limiting clinical acceptability. Although deep learning methods have been used to enhance non-ideal measurement CT images, their reliance on large training datasets and limited generalizability across diverse settings hinder practical use. We propose the multi-scale integrated Transformer Amplifier (TAMP), an imaging foundation model for universal non-ideal measurement CT enhancement. Pre-trained on 10.8 million physics-driven simulated non-ideal measurement CT images, TAMP generalizes effectively across various non-ideal measurement CT settings, defect degrees, and body regions. Moreover, a parameter-efficient fine-tuning strategy enables TAMP to adapt to specific clinical scenarios using only few slices. Extensive experiments, including radiologists and real-world validations, demonstrate that TAMP consistently improves image quality and clinical acceptability, underscoring its significant potential to advance CT imaging and broaden non-ideal measurement CT applications in clinical practice.

Lesion segmentation of acute ischemic stroke (AIS) patients on Non-Contrast CT (NCCT) scans plays a crucial role in rapid diagnosis and treatment planning. The direct application of promising semi-automatic SAM-Med2D to AIS lesion segmentation on NCCT scans: 1) does not achieve good performance; 2) cannot incorporate 3D volumetric context; and 3) requires time-consuming prompts and lacks effective domain-specific semantic prompts. Therefore, we propose a novel SAM-Med2D-based model (AISCT-SAM) with 3D context awareness and self-prompt generation for automatic AIS lesion segmentation. First, to better handle the high heterogeneity of AIS lesions between different subjects, we propose the adapter and low-rank adaptation with gate layer which updates parameters on both aspects of the token-mixer and channel-mixer for more comprehensive and flexible fine-tuning. Second, we introduce a plug-and-play depth adapter to extract richer 3D contextual information, while enhancing the local priors that the Transformer architecture lacks. Then, to enable fully automatic segmentation, we propose a self-prompt generator that generates multi-level semantic prompts by leveraging bilateral hemisphere differences. Finally, we devise a multi-level semantic prompt-guided mask decoder that fully takes into account the characteristics of NCCT images. AISCT-SAM was evaluated on the public AISD dataset, a private AIS dataset, and an external AIS dataset. Experimental results show that AISCT-SAM achieves Dice scores of 63.32%, 48.75%, and 43.81% on the three datasets, respectively, surpassing 20 state-of-the-art methods. Additionally, volumetric analysis and external validation suggest that AISCT-SAM could provide valuable reference information for AIS diagnosis and treatment decisions. Furthermore, AISCT-SAM also performs well on a public MRI dataset, indicating its potential generalization capability for other modalities.

Click-based interactive segmentation aims to delineate a target object with minimal user effort by progressively refining the mask from sparse clicks. Although many recent methods feed the previous prediction (probability map) back into the network, the probability representation is often treated as a weak auxiliary cue, and semantic consistency with the evolving feature representation is rarely enforced. This mismatch tends to accumulate across iterations, introducing redundant signals and weakening refinement in complex scenes, which ultimately demands extra corrective clicks. To address this, we propose DRSM, a Dual-Representation Collaborative Modulation (DRCM) that promotes deep cross-iteration collaboration between probability and feature maps via region-focused refinement and explicit representation alignment. Specifically, DRSM performs bidirectional semantic alignment to stabilize cross-iteration fusion, while a density-driven Region-Focused Strategy (RFS) emphasizes user-indicated areas and suppresses irrelevant interference. Building on this guidance, a Dual-Modulation Enhancement (DME) scheme jointly calibrates the previous probability map and the current multi-scale features, yielding more reliable refinement under sparse interactions. Extensive experiments on five public benchmarks and three out-of-domain datasets show that DRSM achieves accurate and reliable segmentation with fewer clicks, while generalizing robustly to unseen domains.

The inherent spectral properties of photon-counting computed tomography (PCCT) allow detailed material identification through decomposition techniques, but these methods often amplify image noise and artifacts. Current denoising approaches mainly focus on improving already degraded images, ignoring the fundamental noise caused by random variations in photon detection. To tackle these issues, we combine a physics-based noise analysis with deep learning to control noise during the material decomposition process. Our work has three key parts: (1) A noise analysis model that explains how random photon-count variations in the detector affect the noise levels in different materials after decomposition. This model connects the Poisson-distributed detector noise to material-specific noise patterns. (2) A self-supervised training method that combines the noise model with neural networks using probability-based optimization, allowing the system to learn from limited training data without needing high-quality data. (3) A flexible image improvement system that adapts to different body structures and noise conditions, ensuring reliable results across various scanning scenarios. Tests using real patient scan data show our method better preserves material accuracy and produces cleaner virtual monochromatic images compared to traditional approaches. Importantly, our solution works effectively with small training datasets and can be practically used in hospital settings without slowing down workflows. This research bridges the gap between theoretical noise analysis and clinical medical imaging needs, offering a balanced approach to improving PCCT technology.

Low-dose X-ray fluoroscopy is essential in interventional procedures to reduce radiation exposure, yet image quality is inevitably degraded by predominantly Poisson noise, compromising diagnostic reliability. Existing denoising methods commonly rely on large annotated medical datasets, which are scarce in practice, and struggle to balance noise suppression with the preservation of fine structural details. To overcome the above-mentioned problems, we propose SIX, a zero-shot self-supervised interactive denoising framework for low-dose fluoroscopy. The framework integrates a physics-driven noise model with a deep image prior. It performs prior-based denoising via probabilistic modeling with an intuitive hyperparameter, allowing a flexible balance between noise suppression and structural fidelity according to specific diagnostic requirements. Experiments on an in-house dataset show that SIX outperforms various conventional and deep learning methods in PSNR and SSIM, as well as in clinically critical metrics including contrast-to-noise ratio and spatial resolution. With its flexibility and user-friendly design, SIX holds strong potential for deployment in clinical image-guided workflows.

Coronary computed tomography angiography (CCTA) is a pivotal non-invasive imaging modality for diagnosing cardiac disease. However, due to the temporal resolution limitations, cardiac structures, specifically coronary arteries, may suffer from motion artifacts when CCTA is applied to patients with arrhythmias or high heart rates. Limited-angle CT (LA-CT) emerges as a promising alternative by significantly reducing the acquisition time, thereby mitigating the motion artifacts. Yet, LA-CT unavoidably leads to severe wedge artifacts, posing a significant challenge. Therefore, to reconstruct motion-free cardiac CT images while suppressing the wedge artifacts, we propose a Segmentation-Guided Accelerating Diffusion Model (SGADM) tailored for LA-CT imaging. While diffusion models have demonstrated exceptional performance in medical imaging, their extensive sampling procedures impose high computational costs, hindering clinical applicability. To address this issue, SGADM employs an innovative diffusion model that directly generates high-quality CT images. Moreover, SGADM adopts the diffusion perceptual loss to ensure data distribution consistency between two successive sampling steps. As a result, SGADM can provide satisfactory results in fewer than 10 steps. Additionally, SGADM incorporates segmentation guidance to enhance the spatial-positional accuracy of generated coronary arteries explicitly. Both quantitative and qualitative evaluations on simulated and real datasets reveal that the SGADM effectively restores high-quality CCTA images with minimal motion artifacts, highlighting its potential for clinical applications.

Cone Beam Computed Tomography (CBCT) is a real-time imaging modality that provides three-dimensional (3D) images with high-resolution and dose-efficiency. Considering the spatial limitations and the imaging speed constraints, limited-angle sampling scheme is inevitably adopted in several clinical settings. However, limited-angle CBCT (LA-CBCT) reconstruction is a typical ill-posed inverse problem due to insufficient sampling, giving rise to wedge artifacts and structural distortions in the reconstructed images. Although supervised methods have provided possible solutions, these methods require paired LA-CBCT and full-angle CBCT (FA-CBCT) images for training an image restoration model, which are usually unavailable in real-world clinical scenarios. In order to obtain high-quality reconstructed images, this study proposes an unsupervised method named Faithful and Data-Adaptive Reconstruction (FDA-Recon) for LA-CBCT. FDA-Recon comprises three sequential stages, namely Target-to-Source Domain Adaptation (T2SDA) stage, Source Domain Image Restoration (SDIR) stage, and Prior Image Constrained Target Domain Faithful Reconstruction (PIC-TDFR) stage. The T2SDA stage transfers target domain images to source-style images, and then the SDIR stage eliminates artifacts and restores structural information in the transferred images. Finally, the PIC-TDFR stage utilizes an iterative algorithm to minimize domain discrepancy and further enhance the image quality by forcing the consistency to the target domain projections. During the T2SDA and SDIR stages, deep learning (DL) methods are applied. Considering the characteristics of CBCT images, spiral mamba is introduced to enhance the capacity of DL models. Experiments on the real head & neck dataset exhibit the promising performance of FDA-Recon in artifact reduction and image restoration.