Shortening acquisition time in medical imaging, particularly in Magnetic Resonance Imaging (MRI), has been a long-standing challenge involving a trade-off with image quality. This article provides an overview of “deep unfolding,” a technology emerging as a powerful solution to this problem. Deep unfolding is a hybrid approach that fuses classical iterative reconstruction algorithms based on physical models with deep learning that learns features from data, thereby achieving both high performance and interpretability. This paper first explains its fundamental principles and then traces its evolution in MRI reconstruction, from pioneering models like ADMM-Net and ISTA-Net, to advanced architectures such as MoDL that integrate powerful priors like U-Net, and finally to the latest architectures applying Transformers. Next, we discuss the trend of self-supervised learning, which overcomes the need for fully-sampled ground-truth data―a major barrier to clinical application. Furthermore, a theoretical interpretation of the behavior of learned parameters is introduced to explain why deep unfolding achieves rapid convergence. Finally, as a fundamental challenge for the technology’s widespread clinical adoption, we address the issue of generalization to out-of-distribution data and provide a perspective on future research directions.

One approach to image reconstruction from observed data is to minimize an objective function composed of a data fidelity term based on the observation model and a regularization term reflecting the properties of the original images. While various regularization functions for image restoration have been manually designed, recent studies have proposed data-driven methods that employ regularizers learned from data. This article provides an overview of such approaches based on learned regularization functions and their characteristics. Furthermore, we present simulation results that demonstrate the application of learned regularizers to computed tomography (CT) image reconstruction.

X-ray ptychography is a technique that visualizes samples with spatial resolutions on the order of tens of nanometers by phase retrieval from diffraction intensity patterns obtained through X-ray illumination. Since imaging quality strongly depends on phase retrieval, it is essential to develop phase retrieval methods that are robust to noise and adaptable to diverse experimental conditions. In recent years, many phase retrieval approaches based on deep learning have been proposed; however, they face challenges such as the difficulty of preparing large training datasets and limited adaptability to various experimental conditions. A promising solution to these issues is the fusion of deep learning and optimization, an approach that has recently attracted considerable attention in the fields of signal processing, machine learning, and optimization. In this article, we introduce phase retrieval methods based on this hybrid approach and explain how they achieve both flexibility to changing measurement conditions and high reconstruction performance.

In this study, we present a novel framework for accelerated MRI reconstruction that introduces posterior probability sampling based on diffusion models within a divide-and-conquer strategy. In accelerated MRI, patient anatomy must be reconstructed from under-sampled k-space measurements, which makes the problem inherently ill-posed. To address this problem, prior knowledge is incorporated to restrict the solution space and improve the fidelity of the reconstructed images. Specifically, we employ a prior probability distribution modeled by a diffusion process to constrain reconstruction. For each posterior sample, the Fourier transform image is analyzed, and reconstruction accuracy is estimated as a confidence value for each frequency component. By sequentially adopting frequency sequences with higher confidence, the method progressively enhances reconstruction quality while gradually tightening constraints. Furthermore, by generating multiple reconstructions through repeated posterior sampling, we compute the pixel-wise variance of the reconstructed images. This variance serves as a statistical indicator of reconstruction reliability, allowing the confidence of individual pixels to be visualized and quantified. The proposed framework thus provides both accurate image reconstruction and an interpretable measure of uncertainty. Finally, we demonstrate the effectiveness of our method through experiments on real MRI data.

This survey synthesizes evidence on spectral beam shaping in X-ray computed tomography (CT) using tin/silver filtration (Sn/Ag-filtered acquisition), focusing on clinical applications, radiation dose, and image quality. Target domains include chest imaging, pediatrics, paranasal sinuses, musculoskeletal examinations, urinary stone detection, coronary calcium scoring, interventional radiology, and the localizer radiograph. Across indications, many studies reported substantial reductions in output dose and effective dose (approximately 20-90％) with preserved diagnostic performance. Trade-offs include decreased tissue contrast in some organs, limitations on applicable tube voltages, and dependence on patient diameters and scanner output. Sn/Ag filtration increases the effective beam energy by approximately 20-30％. Overall, spectral beam shaping is useful option for non-contrast, high-contrast tasks (e.g., chest and bone) and scenarios requiring repeated scanning.

In recent years, improvements in GPU performance have made it possible to construct deep learning models for voxel data. However, chest CT images typically have a resolution of around 200-300×512×512 per case, and processing data of this scale requires substantial GPU memory and extensive training time. Therefore, we propose methods for the effective utilization of voxel data in chest CT images. The first method uses data obtained by sampling slices based on the three axes: the coronal axis, the sagittal axis, and the axial axis. The classification performance for the minority class was evaluated using precision-recall area under the curve (PR-AUC)，and the results showed a mean accuracy of 0.9948 with a standard deviation of 0.0103 for data in the same domain as the training data and a mean of 0.5177 with a standard deviation of 0.1320 for data in different domains. The second method extracts brightness and texture features from the images and inputs the concatenated data into the model. This method maintained high classification performance for data in the same domain as the training data with a mean of 0.9071 and a standard deviation of 0.1008, while also achieving high classification performance with a mean of 0.8174 and a standard deviation of 0.0546 for data in different domains.