LiMT: A Multi-task Liver Image Benchmark Dataset

A new framework named PETRG-3D has been proposed for automated 3D PET/CT report generation, addressing the growing need for efficient reporting in oncology due to a shortage of trained specialists. This model utilizes a dual-branch architecture to separately encode PET and CT volumes while incorporating style-adaptive prompts to standardize reporting across different hospitals.

MedSAM-3 has been introduced as a text promptable medical segmentation model designed to enhance medical image and video segmentation by allowing precise targeting of anatomical structures through open-vocabulary text descriptions. This model builds on the Segment Anything Model (SAM) 3 architecture, addressing the limitations of existing methods that require extensive manual annotation for clinical applications.

A novel multi-view X-ray/CT registration method has been proposed to enhance intraoperative bone registration in orthopedic procedures, addressing challenges such as achieving sub-millimeter accuracy and reducing the need for manual key-point annotations. This method utilizes contour-based iterative closest point (ICP) optimization, focusing on specific bone substructure contours to improve registration robustness and accuracy.

A new framework named ReBrain has been introduced for reconstructing brain MRI from sparse CT slices using a retrieval-augmented diffusion approach. This method employs a Brownian Bridge Diffusion Model to synthesize MRI slices and retrieves similar CT slices from a prior database to enhance reconstruction accuracy. This innovation addresses the challenges posed by low-dose CT scans that often result in poor resolution and limited data for MRI synthesis.

A new model named FCDM has been introduced, designed specifically for sinogram inpainting in computed tomography (CT). This model addresses the challenges posed by sparse-view CT, which often results in incomplete sinograms and structured signal loss, hindering accurate image reconstruction. FCDM utilizes a diffusion-based framework that incorporates bidirectional frequency reasoning and angular-aware masking to restore the global structure of sinograms while adhering to physical constraints.

A novel approach to linear inverse problems has been introduced through an unconditional representation of the conditional score function (UCoS), which significantly reduces computational costs by eliminating the need for multiple forward model evaluations during sampling. This method leverages an offline training phase to learn a task-dependent score function based on the linear forward operator.