Dong Liang

Purpose: Masked modeling has emerged as a powerful deep learning paradigm that reconstructs missing data by predicting masked regions, achieving strong performance across multiple domains. As a critical task in medical imaging, magnetic resonance imaging (MRI) reconstruction aims to recover high‐quality images from under‐sampled k‐space data. However, existing methods usually optimize the full image domain or k‐space, while overlooking the importance and structural roles of different frequency components. Methods: An adaptive mask‐guided diffusion model (AMDM) was developed, which leverages adaptively estimated frequency distributions from k‐space to construct a hybrid mask mechanism for handling diverse k‐space inputs. It enables the effective separation of high‐frequency and low‐frequency components, producing multiple prior representations. In addition, the frequency distribution is used to generate adaptive masks, which subsequently guide the closed‐loop diffusion process. Results: Experimental results demonstrated improved reconstruction performance through frequency‐aware modeling. AMDM achieved a PSNR of 36.89 dB on an independent public brain dataset without retraining, even at an under‐sampling factor of . Conclusion: The proposed method shows that explicitly modeling k‐space frequency heterogeneity provides an effective prior for MRI reconstruction, offering a promising direction for robust reconstruction under aggressive under‐sampling.

The Segment Anything Model (SAM) has demonstrated groundbreaking performance in natural image segmentation, yet its direct application to medical imaging remains suboptimal due to domain shifts in data distributions and the inherent 3D nature of medical data. Although recent SAM-based methods have employed parameter-efficient transfer learning (PETL) to adapt SAM for medical tasks, they often overlook the critical 3D contextual information essential for accurate volumetric segmentation. To address this limitation, we introduce VSS-SAM++, a novel dual-branch architecture that combines SAM's foundational visual priors with Vision Mamba's capacity for modeling long-range spatial dependencies. In this framework, SAM serves as the primary encoder for high-level feature extraction, while a parallel Mamba branch captures cross-slice dependencies in 3D medical volumes. A gated hybrid attention module then dynamically fuses complementary features from both branches, adaptively weighting multi-view representations to minimize feature ambiguity and enhance segmentation precision. Extensive evaluations across nine public CT and MRI datasets demonstrate that VSS-SAM++ outperforms existing methods by 0.2-11.3% in Dice score on multi-organ and lesion segmentation tasks. The framework's robustness to domain shifts and scalability across diverse modalities highlights its potential for clinical deployment.

Purpose: To propose ESPIRiT‐Diffusion, a physics‐guided score‐based diffusion reconstruction framework incorporating multi‐set ESPIRiT map‐based data‐consistency constraints for 8.8‐ and 10.7‐fold accelerated joint intracranial and carotid vessel wall imaging (VWI) with an isotropic resolution of 0.6 mm ³ . Theory and Methods: ESPIRiT‐Diffusion exploits the powerful generative capability of the diffusion framework for the reconstruction of large‐FOV 3D VWI images, aiming to recover vessel wall details and improve image quality at high acceleration factors. By further incorporating multi‐set ESPIRiT coil sensitivity maps into the Langevin equation, it enforces accurate data consistency, thereby improving VWI reconstruction quality while constraining unreliable generation. In addition, diffusion performed directly in the image domain leads to clearer fine details and faster reconstruction. Results: In retrospective experiments with Cartesian, CAIPI, and variable‐density undersampling at acceleration factors of 8.8× and 10.7×, ESPIRiT‐Diffusion showed improved reconstruction performance compared with ESPIRiT, DL‐ESPIRiT, SENSE‐Diffusion, and SPIRiT‐Diffusion in the evaluated retrospective experiments, with better preservation of fine vessel wall structures. In prospective patient experiments, ESPIRiT‐Diffusion provided favorable visualization of vessel wall lesions, with no statistically significant differences in reader scores from the 3‐fold CS reference across either individual vascular segments or Overall comparisons. Conclusions: ESPIRiT‐Diffusion for VWI reconstruction mitigates some limitations related to instability and sampling‐pattern dependence in unfolding‐based methods, while also alleviating image blurring and reducing reconstruction time compared with k‐space diffusion. As a result, ESPIRiT‐Diffusion showed improved reconstruction quality and clearer fine structural details in the evaluated experiments, while reducing the required acquisition time.

Existing learning-based visual retouching primarily focuses on improving image quality through end-to-end objective mapping between input and retouched images. However, these approaches often overlook two critical aspects: the progressive nature of image retouching and the subjective aesthetic preferences, resulting in suboptimal visual outcomes. To address this, we introduce Automatic Aesthetic Image Retouching via active reinforcement learning (A$^{3}$3RL) to enhance the visualization experience in two sub-tasks: color enhancement and composition optimization, which are formulated as a unified Markov Decision Process in the proposed A$^{3}$3RL framework. In our approach, each pixel functions as an autonomous agent that determines optimal actions based on aesthetic guidance, engaging in online exploration through immediate pixel-wise and channel-wise feedback from the aesthetic environment. By leveraging a pretrained image aesthetic model, our method ensures that the A$^{3}$3RL process aligns with human aesthetic preferences and adheres to subjective aesthetic principles. The framework integrates pixel-level retouching actions with image-level operations to achieve optimal image sequences through progressive iterations. Extensive experiments demonstrate that our method effectively recalibrates image aesthetics across multiple dimensions: low-level quality metrics (PSNR, SSIM), visual perception (LPIPS), and subjective visual experience (human survey). The results demonstrate high consistency with expert-retouched ground-truth images.

Whole-body ¹⁸ F-fluorodeoxyglucose positron emission tomography/computed tomography ( ¹⁸ F-FDG PET/CT) is central to lung cancer staging, yet report drafting remains time intensive and error sensitive, as a single mislocalized finding can alter TNM stage and treatment decisions. We cast PET/CT report drafting as TNM-accountable decision support and propose RIDE, a 2-stage dual-modality 3-dimensional framework that separates TNM-critical anchoring from narrative completion. Stage I generates TNM-oriented structured impressions from paired PET and CT volumes; stage II performs hierarchical organ-wise exemplar retrieval and prompts a large language model to synthesize complete Findings and Impression conditioned on these inspectable artifacts. We curate a multicenter whole-body FDG PET/CT cohort of 1,583 patients from 3 hospitals, all referred for suspected or confirmed lung cancer, and evaluate it on a held-out test set of 520 cases, including 276 external cases from 2 independent institutions. RIDE achieves the strongest overall drafting performance and cross-site generalization, improving a clinically grounded competency metric by +20.4 on the external set and maintaining high clinician ratings under shift (mean Likert 4.00 internal; 4.17 external) while also delivering the best TNM staging performance among evaluated baselines. These findings support workflow-decomposed, TNM-accountable PET/CT drafting as a promising decision-support approach for human verification, and prospective studies will be valuable for further characterizing its utility in real-world clinical workflows.