Image Pairs Research Articles

Intrafractional motion detection for spine SBRT via X-ray imaging using ExacTrac Dynamic

PurposeDue to its close vicinity to critical structures, especially the spinal cord, standards for safety for spine stereotactic body radiotherapy (SBRT) should be high. This study was conducted, to evaluate intrafractional motion during spine SBRT for patients without individualized immobilization (e.g., vacuum cushions) using high accuracy patient monitoring via orthogonal X-ray imaging. MethodsIntrafractional X-ray data were collected from 29 patients receiving 79 fractions of spine SBRT. No individualized immobilization devices were used during the treatment. Intrafractional motion was monitored using the ExacTrac Dynamic (ETD) System (Brainlab AG, Munich, Germany). Deviations were detected in six degrees of freedom (6 DOF). Tolerances for repositioning were 0.7 mm for translational and 0.5° for rotational deviations. Patients were repositioned when the tolerance levels were exceeded. ResultsOut of the 925 pairs of stereoscopic X-ray images examined, 138 (15 %) showed at least one deviation exceeding the predefined tolerance values. In all 6 DOF together, a total of 191 deviations out of tolerance were recorded. The frequency of deviations exceeding the tolerance levels varied among patients but occurred in all but one patient. Deviations out of tolerance could be seen in all 6 DOF. Maximum translational deviations were 2.6 mm, 2.3 mm and 2.8 mm in the lateral, longitudinal and vertical direction. Maximum rotational deviations were 1.8°, 2.6° and 1.6° for pitch, roll and yaw, respectively. Translational deviations were more frequent than rotational ones, and frequency and magnitude of deviations showed an inverse correlation. ConclusionIntrafractional motion detection and patient repositioning during spine SBRT using X-ray imaging via the ETD System can lead to improved safety during the application of high BED in critical locations. When using intrafractional imaging with low thresholds for re-positioning individualized immobilization devices (e.g. vacuum cushions) may be omitted.

A comprehensive lung CT landmark pair dataset for evaluating deformable image registration algorithms.

Deformable image registration (DIR) is a key enabling technology in many diagnostic and therapeutic tasks, but often does not meet the required robustness and accuracy for supporting clinical tasks. This is in large part due to a lack of high-quality benchmark datasets by which new DIR algorithms can be evaluated. Our team was supported by the National Institute of Biomedical Imaging and Bioengineering to develop DIR benchmark dataset libraries for multiple anatomical sites, comprising of large numbers of highly accurate landmark pairs on matching blood vessel bifurcations. Here we introduce our lung CT DIR benchmark dataset library, which was developed to improve upon the number and distribution of landmark pairs in current public lung CT benchmark datasets. Thirty CT image pairs were acquired from several publicly available repositories as well as authors' institution with IRB approval. The data processing workflow included multiple steps: (1) The images were denoised. (2) Lungs, airways, and blood vessels were automatically segmented. (3) Bifurcations were directly detected on the skeleton of the segmented vessel tree. (4) Falsely identified bifurcations were filtered out using manually defined rules. (5) A DIR was used to project landmarks detected on the first image onto the second image of the image pair to form landmark pairs. (6) Landmark pairs were manually verified. This workflow resulted in an average of 1262 landmark pairs per image pair. Estimates of the landmark pair target registration error (TRE) using digital phantoms were 0.4mm±0.3mm. The data is published in Zenodo at https://doi.org/10.5281/zenodo.8200423. Instructions for use can be found at https://github.com/deshanyang/Lung-DIR-QA. The dataset library generated in this work is the largest of its kind to date and will provide researchers with a new and improved set of ground truth benchmarks for quantitatively validating DIR algorithms within the lung.

Adapting low-dose CT denoisers for texture preservation using zero-shot local noise-level matching.

On enhancing the image quality of low-dose computed tomography (LDCT), various denoising methods have achieved meaningful improvements. However, they commonly produce over-smoothed results; the denoised images tend to be more blurred than the normal-dose targets (NDCTs). Furthermore, many recent denoising methods employ deep learning(DL)-based models, which require a vast amount of CT images (or image pairs). Our goal is to address the problem of over-smoothed results and design an algorithm that works regardless of the need for a large amount of training dataset to achieve plausible denoising results. Over-smoothed images negatively affect the diagnosis and treatment since radiologists had developed clinical experiences with NDCT. Besides, a large-scale training dataset is often not available in clinical situations. To overcome these limitations, we propose locally-adaptive noise-level matching (LANCH), emphasizing the output should retain the same noise-level and characteristics to that of the NDCT without additionaltraining. We represent the NDCT image as the pixel-wisely weighted sum of an over-smoothed output from off-the-shelf denoiser (OSD) and the difference between the LDCT image and the OSD output. Herein, LANCH determines a 2D ratio map (i.e., pixel-wise weight matrix) by locally matching the noise-level of output and NDCT, where the LDCT-to-NDCT device flux (mAs) ratio reveals the NDCT noise-level. Thereby, LANCH can preserve important details in LDCT, and enhance the sharpness of the noise-free regions. Note that LANCH can enhance any LDCT denoisers without additional training data (i.e., zero-shot). The proposed method is applicable to any OSD denoisers, reporting significant texture plausibility development over the baseline denoisers in quantitative and qualitative manners. It is surprising that the denoising accuracy achieved by our method with zero-shot denoiser was comparable or superior to that of the best training-based denoisers; our result showed 1% and 33% gains in terms of SSIM and DISTS, respectively. Reader study with experienced radiologists shows significant image quality improvements, a gain of +1.18 on a five-point mean opinion scorescale. In this paper, we propose a technique to enhance any low-dose CT denoiser by leveraging the fundamental physical relationship between the x-ray flux and noise variance. Our method is capable of operating in a zero-shot condition, which means that only a single low-dose CT image is required for the enhancement process. We demonstrate that our approach is comparable or even superior to supervised DL-based denoisers that are trained using numerous CT images. Extensive experiments illustrate that our method consistently improves the performance of all tested LDCT denoisers.

Rapid and Label-Free Histopathology of Oral Lesions Using Deep Learning Applied to Optical and Infrared Spectroscopic Imaging Data.

Oral potentially malignant disorders (OPMDs) are precursors to over 80% of oral cancers. Hematoxylin and eosin (H&E) staining, followed by pathologist interpretation of tissue and cellular morphology, is the current gold standard for diagnosis. However, this method is qualitative, can result in errors during the multi-step diagnostic process, and results may have significant inter-observer variability. Chemical imaging (CI) offers a promising alternative, wherein label-free imaging is used to record both the morphology and the composition of tissue and artificial intelligence (AI) is used to objectively assign histologic information. Here, we employ quantum cascade laser (QCL)-based discrete frequency infrared (DFIR) chemical imaging to record data from oral tissues. In this proof-of-concept study, we focused on achieving tissue segmentation into three classes (connective tissue, dysplastic epithelium, and normal epithelium) using a convolutional neural network (CNN) applied to three bands of label-free DFIR data with paired darkfield visible imaging. Using pathologist-annotated H&E images as the ground truth, we demonstrate results that are 94.5% accurate with the ground truth using combined information from IR and darkfield microscopy in a deep learning framework. This chemical-imaging-based workflow for OPMD classification has the potential to enhance the efficiency and accuracy of clinical oral precancer diagnosis.

The One-Point-One-Line geometry for robust and efficient line segment correspondence

Three-dimensional (3D) lines are common elements in artificial scenes and serve as basic, yet essential features for structural 3D reconstruction. The crucial step of 3D line reconstruction, namely two-view line segment matching, still faces challenges in terms of both accuracy and efficiency improvements. Therefore, robust and efficient constraints are needed to establish valid line candidates. This paper introduces a novel geometry constraint called “one-point-one-line geometry” (OPOL) to enhance the precision of line matching and reduce computational complexity. OPOL offers two remarkable advantages: (1) It takes point orientations as the constraint, which is not only invariant to projective transformations, but also alleviates computational requirements. (2) It needs only one point match to construct the geometry constraint, thus both the grouping and validation are greatly reduced. Additionally, we incorporate the line sweep strategy into OPOL, leveraging depth and space constraints derived from existing 3D points to further enhance efficiency. Extensive experiments conducted on large-coverage and high-resolution images (as large as 10336 × 7788 pixels) demonstrated that OPOL matched lines within a second for an image pair. Both quantitative and qualitative results also demonstrated the superior accuracy and efficiency performance of OPOL. We integrated OPOL into multiple view line reconstruction frameworks, and the promising experimental results reveal the performance of OPOL for robust line reconstruction. The OPOL code is publicly available at https://github.com/JoeAlexxxxx/OPOL.

WISE: whole-scenario embryo identification using self-supervised learning encoder in IVF.

To study the effectiveness of whole-scenario embryo identification using a self-supervised learning encoder (WISE) in in vitro fertilization (IVF) on time-lapse, cross-device, and cryo-thawed scenarios. WISE was based on the vision transformer (ViT) architecture and masked autoencoders (MAE), a self-supervised learning (SSL) method. To train WISE, we prepared three datasets including the SSL pre-training dataset, the time-lapse identification dataset, and the cross-device identification dataset. To identify whether pairs of images were from the same embryos in different scenarios in the downstream identification tasks, embryo images including time-lapse and microscope images were first pre-processed through object detection, cropping, padding, and resizing, and then fed into WISE to get predictions. WISE could accurately identify embryos in the three scenarios. The accuracy was 99.89% on the time-lapse identification dataset, and 83.55% on the cross-device identification dataset. Besides, we subdivided a cryo-thawed evaluation set from the cross-device test set to have a better estimation of how WISE performs in the real-world, and it reached an accuracy of 82.22%. There were approximately 10% improvements in cross-device and cryo-thawed identification tasks after the SSL method was applied. Besides, WISE demonstrated improvements in the accuracy of 9.5%, 12%, and 18% over embryologists in the three scenarios. SSL methods can improve embryo identification accuracy even when dealing with cross-device and cryo-thawed paired images. The study is the first to apply SSL in embryo identification, and the results show the promise of WISE for future application in embryo witnessing.

MRFormer: Multiscale retractable transformer for medical image progressive denoising via noise level estimation

Clear medical images are important for auxiliary diagnoses, but the images generated by various medical devices inevitably contain considerable noise. Although various models have been proposed for denoising, these methods ignore the fact that different types of medical images have different noise levels, which leads to unsatisfactory test results. In addition, collecting many medical images for training denoising models consumes many material resources. To address these issues, we formulate a progressive denoising architecture that contains preliminary and profound denoising. First, we construct a noise level estimation network to estimate the noise level via self-supervised learning and perform preliminary denoising with a dilated blind-spot network. Second, with the learned noise distribution, we synthesize noisy natural images to construct clean-noisy natural image pairs. Finally, we design a novel medical image denoising model for profound denoising by training these pairs. The proposed three-stage learning scheme and progressive denoising architecture not only solve the problem that the denoising model only adapts to a single noise level but also alleviate the lack of medical image pairs. Moreover, we integrate dense attention and sparse attention to constitute the retractable transformer module in the profound denoising model, which reconciles a wider receptive field and enhances the representation ability of the transformer, s allowing the denoising model to obtain retractable attention on the input feature and capture more local and global receptive fields simultaneously. The results of qualitative and quantitative experiments demonstrate the effectiveness of our method in removing noise at various levels.

DMHomo: Learning Homography with Diffusion Models

Supervised homography estimation methods face a challenge due to the lack of adequate labeled training data. To address this issue, we propose DMHomo, a diffusion model-based framework for supervised homography learning. This framework generates image pairs with accurate labels, realistic image content, and realistic interval motion, ensuring they satisfy adequate pairs. We utilize unlabeled image pairs with pseudo-labels such as homography and dominant plane masks, computed from existing methods, to train a diffusion model that generates a supervised training dataset. To further enhance performance, we introduce a new probabilistic mask loss, which identifies outlier regions through supervised training, and an iterative mechanism to optimize the generative and homography models successively. Our experimental results demonstrate that DMHomo effectively overcomes the scarcity of qualified datasets in supervised homography learning and improves generalization to real-world scenes. The code and dataset are available at: https://github.com/lhaippp/DMHomo

Coupled image and kernel prior learning for high-generalized super-resolution

Deep convolution neural networks (DCNNs) have demonstrated great success on single image super resolution, where most existing methods aim to construct a deep model in an fully-supervised way using large amount of synthetic image pairs under ideal degradation assumption. The super-resolution performance would be significantly degraded for the low-resolution images captured under uncontrolled imaging conditions with complicated degradation procedures. To handle the above limitations, this work proposes a high-generalized image super-resolution framework by synergistically learning the latent image and the degradation process to achieve the specific priors for an under-studying observation in an unsupervised manner. Specifically, we incorporate dual branches of networks to configure our framework, where one is structured with both convolution and transformer blocks to learn local-to-global prior to generate high-quality image while the other aims to learn the kernel prior with a simple convolution-based architecture. The kernel prior subnet is pre-trained to enhance the stability of the joint optimization for the overall unsupervised learning procedure. With the estimated degradation kernel and target image, we produce the approximated low-resolution version with a convolution-based degradation block to formulate the loss function for the specific learning. Moreover, to pursue visually plausible image generation, we further incorporate a perceptual loss by leveraging a pre-trained discriminator in Gan-based SR model. Extensive experiments on several benchmark datasets have demonstrated the effectiveness of our proposed high-generalized super-resolution model, manifesting superiority over both supervised and unsupervised state-of-the-art SR models.

A new approach to medical image fusion based on the improved Extended difference-of-Gaussians combined with the Coati optimization algorithm

The synthesis of medical images plays a pivotal role in image-based disease diagnosis. In recent years, numerous medical image synthesis methods have been proposed. Nevertheless, images generated from the proposed synthesis methods often suffer from shortcomings, including low image quality, reduced brightness and contrast, and loss of vital information. In this paper, we propose a novel approach to tackle the aforementioned challenges in medical image synthesis. Initially, the input images are decomposed into two components: low-frequency and high-frequency components using the Weighted mean curvature filter (WMCF). Subsequently, we propose a synthesis rule for the high-frequency components based on the combination of the Extended difference-of-Gaussians (XDoG) filter, the Structure tensor (ST), and the Local energy (LE) function. Additionally, we employ a novel adaptive synthesis rule, based on the Coati optimization algorithm (COA), to synthesize the low-frequency components. We conducted four experiments using 90 pairs of medical images. The experimental results demonstrate that our proposed method not only effectively enhances image quality, brightness, and contrast but also better preserves crucial details such as boundaries, edges, and the original image’s structure when compared to the most recently published methods.

Unsupervised remote sensing image thin cloud removal method based on contrastive learning

AbstractCloud removal algorithm is a crucial step of remote sensing image preprocessing. The current mainstream remote sensing image cloud removal algorithms are implemented based on deep learning, and most of them are supervised. A large number of data pairs are required for training to achieve cloud removal. However, real with/without cloud image pairs datasets are difficult to obtain in the real world, and the models obtained by training on synthetic datasets often need to generalize better to natural scenes. And the existing unsupervised thin cloud removal methods based on Cycle‐GAN framework with considerable model complexity and unstable training are not an excellent solution to the problem of lack of paired datasets. Based on this, in this paper, the authors propose an unsupervised remote sensing image thin cloud removal method based on contrastive learning—GAN‐UD. It is a network consisting of a frequency‐spatial attention generator and a discriminator. In addition, the authors introduce local contrastive loss and global content loss to constrain the content of the generated images to ensure that the generated cloud‐free images are consistent with the input cloud images in terms of image content. Experimental results show that the proposed method in this paper can still effectively remove thin clouds from remote sensing images without paired training datasets, outperforms current unsupervised cloud removal methods, and achieves comparable performance to supervised methods.

DSCNet: lightweight and efficient self-supervised network via depthwise separable cross convolution blocks for speckle image matching.

Speckle structured light has become a research hotspot due to its ability to acquire target three-dimensional information with single image projection in recent years. To address the challenges of a low number of extracted speckle feature points, high mismatch rate and poor real-time performance in traditional algorithms, as well as the obstacle of requiring expensive annotation data in deep learning-based methods, a lightweight and efficient self-supervised convolutional neural network (CNN) is proposed to achieve high-precision and rapid matching of speckle images. First, to efficiently utilize the speckle projection information, a feature extraction backbone based on the depthwise separable cross convolution blocks is proposed. Second, in the feature detection module, a softargmax detection head is designed to refine the coordinates of speckle feature points to sub-pixel accuracy. In the feature description module, a coarse-to-fine module is presented to further refine matching accuracy. Third, we adopt strategies of transfer learning and self-supervised learning to improve the generalization and feature representation capabilities of the model. Data augmentation and real-time training techniques are used to improve the robustness of the model. The experimental results show that the proposed method achieves a mean matching accuracy of 91.62% for speckle feature points on the pilot's helmet, with mere 0.95% mismatch rate. The full model runs at 42ms for a speckle image pair on an RTX 3060.

ERROR ANALYSIS OF VISUAL ODOMETRY FOR A SMALL SIZE UNMANNED AERIAL VEHICLE

Abstract. Usage of small size Unmanned Aerial Vehicle (UAV) with autonomous flight capability became widespread at both civil and military applications because of its low costs and ease of use. However, inertial or GPS based positioning and orientation sensors can be disabled at battlefields and the UAV losses its path in a short time. Visual odometry becomes a solution for the aforementioned hassle. Real-time detection of ground control points, matching the same control points at the stereo pair images and correcting the position and orientation data with least square adjustment are the challenges of visual odometry. In this study, error analysis of position and orientation of a UAV flying through mountainous region which cannot obtain any augmenting data for position and orientation is conducted by error propagation law. The flight of UAV is simulated with exact initial position and orientation data. Optic images of ground obtained by nadir directed camera are georeferenced using the digital elevation model of the region, interior and exterior camera parameters. The digital elevation model is used to correct the relief displacement. The stereo-pair image is obtained when the UAV flies 200 m and image coordinates of the common control points are detected by colinearity equations. Position and orientation of UAV is corrected by least square adjustment and the process is repeated for each obtained image acquired at 200 m interval. The analysis revealed that the precisions of horizontal and vertical positioning becomes 75 and 200 meter, respectively. The precision of roll angle becomes worse than 0.5 degree, while the precisions of pitch and yaw angle are estimate to be better that 0.1 degree.

TGLFusion: A Temperature-Guided Lightweight Fusion Method for Infrared and Visible Images.

The fusion of infrared and visible images is a well-researched task in computer vision. These fusion methods create fused images replacing the manual observation of single sensor image, often deployed on edge devices for real-time processing. However, there is an issue of information imbalance between infrared and visible images. Existing methods often fail to emphasize temperature and edge texture information, potentially leading to misinterpretations. Moreover, these methods are computationally complex, and challenging for edge device adaptation. This paper proposes a method that calculates the distribution proportion of infrared pixel values, allocating fusion weights to adaptively highlight key information. It introduces a weight allocation mechanism and MobileBlock with a multispectral information complementary module, innovations which strengthened the model's fusion capabilities, made it more lightweight, and ensured information compensation. Training involves a temperature-color-perception loss function, enabling adaptive weight allocation based on image pair information. Experimental results show superiority over mainstream fusion methods, particularly in the electric power equipment scene and publicly available datasets.

Independently Trained Multi-Scale Registration Network Based on Image Pyramid.

Image registration is a fundamental task in various applications of medical image analysis and plays a crucial role in auxiliary diagnosis, treatment, and surgical navigation. However, cardiac image registration is challenging due to the large non-rigid deformation of the heart and the complex anatomical structure. To address this challenge, this paper proposes an independently trained multi-scale registration network based on an image pyramid. By down-sampling the original input image multiple times, we can construct image pyramid pairs, and design a multi-scale registration network using image pyramid pairs of different resolutions as the training set. Using image pairs of different resolutions, train each registration network independently to extract image features from the image pairs at different resolutions. During the testing stage, the large deformation registration is decomposed into a multi-scale registration process. The deformation fields of different resolutions are fused by a step-by-step deformation method, thereby addressing the challenge of directly handling large deformations. Experiments were conducted on the open cardiac dataset ACDC (Automated Cardiac Diagnosis Challenge); the proposed method achieved an average Dice score of 0.828 in the experimental results. Through comparative experiments, it has been demonstrated that the proposed method effectively addressed the challenge of heart image registration and achieved superior registration results for cardiac images.

DCENet-based low-light image enhancement improved by spiking encoding and convLSTM.

The direct utilization of low-light images hinders downstream visual tasks. Traditional low-light image enhancement (LLIE) methods, such as Retinex-based networks, require image pairs. A spiking-coding methodology called intensity-to-latency has been used to gradually acquire the structural characteristics of an image. convLSTM has been used to connect the features. This study introduces a simplified DCENet to achieve unsupervised LLIE as well as the spiking coding mode of a spiking neural network. It also applies the comprehensive coding features of convLSTM to improve the subjective and objective effects of LLIE. In the ablation experiment for the proposed structure, the convLSTM structure was replaced by a convolutional neural network, and the classical CBAM attention was introduced for comparison. Five objective evaluation metrics were compared with nine LLIE methods that currently exhibit strong comprehensive performance, with PSNR, SSIM, MSE, UQI, and VIFP exceeding the second place at 4.4% (0.8%), 3.9% (17.2%), 0% (15%), 0.1% (0.2%), and 4.3% (0.9%) on the LOL and SCIE datasets. Further experiments of the user study in five non-reference datasets were conducted to subjectively evaluate the effects depicted in the images. These experiments verified the remarkable performance of the proposed method.

End-to-End Edge-Guided Multi-Scale Matching Network for Optical Satellite Stereo Image Pairs

End-to-End Edge-Guided Multi-Scale Matching Network for Optical Satellite Stereo Image Pairs

Looking Beyond Single Images for Weakly Supervised Semantic Segmentation Learning.

This article studies the problem of learning weakly supervised semantic segmentation (WSSS) from image-level supervision only. Current popular solutions leverage object localization maps from classifiers as supervision for semantic segmentation learning, and struggle to make the localization maps capture more complete object content. Rather than previous efforts that primarily focus on intra-image information, we address the value of cross-image semantic relations for comprehensive object pattern mining. To achieve this, two neural co-attentions are incorporated into the classifier to complementarily capture cross-image semantic similarities and differences. In particular, given a pair of training images, one co-attention enforces the classifier to recognize the common semantics from co-attentive objects, while the other one, called contrastive co-attention, drives the classifier to identify the unique semantics from the rest, unshared objects. This helps the classifier discover more object patterns and better ground semantics in image regions. In addition to boosting object pattern learning, the co-attention can leverage context from other related images to improve localization map inference, hence eventually benefiting semantic segmentation learning. More importantly, our algorithm provides a unified framework that handles well different WSSS settings, i.e., learning WSSS with 1) precise image-level supervision only, 2) extra simple single-label data, and 3) extra noisy web data. Without bells and whistles, it sets new state-of-the-arts on all these settings. Moreover, our approach ranked 1st place in the Weakly-Supervised Semantic Segmentation Track of CVPR2020 Learning from Imperfect Data Challenge. The extensive experimental results demonstrate well the efficacy and high utility of our method.

MeltpoolGAN: Melt pool prediction from path-level thermal history

In metal laser powder bed fusion, a laser beam rapidly scans through a thin layer of powder. The laser heats up and liquefies the powder forming a melt pool. The melt pool characteristics affect the material property and physical performance of the part. Physical simulations of melt pool dynamics are computationally expensive and are often limited to a short laser scan. Alternatively, Finite Element (FE) simulations for part-scale thermal history are less computationally intensive but do not capture the detailed melt pool temperature profile and geometry in general. Accurate modeling of the high spatial and temporal thermal gradients inside and near the melt pool is essential to many applications of the thermal history including design and optimization of the laser scan path, material microstructure characterization, and deformation prediction.In this work, we developed MeltpoolGAN, the first conditional Generative Adversarial Network (cGAN) to predict the melt pool image based on the simulated thermal history along the laser scan path. The integration of generative deep learning and thermal history simulation achieves melt pool-level accuracy while significantly reducing the physical and computational complexity. The data pair of thermal history and melt pool images for neural network training and validation are obtained through Contact-Aware Path Level (CAPL) thermal simulation and a co-axial melt pool monitoring system developed by NIST, respectively. The predicted melt pool morphology is validated against experimentally acquired melt pool images through geometric characteristics, including the length and width, of the melt pool.

Zero-Shot Medical Image Translation via Frequency-Guided Diffusion Models.

Recently, the diffusion model has emerged as a superior generative model that can produce high quality and realistic images. However, for medical image translation, the existing diffusion models are deficient in accurately retaining structural information since the structure details of source domain images are lost during the forward diffusion process and cannot be fully recovered through learned reverse diffusion, while the integrity of anatomical structures is extremely important in medical images. For instance, errors in image translation may distort, shift, or even remove structures and tumors, leading to incorrect diagnosis and inadequate treatments. Training and conditioning diffusion models using paired source and target images with matching anatomy can help. However, such paired data are very difficult and costly to obtain, and may also reduce the robustness of the developed model to out-of-distribution testing data. We propose a frequency-guided diffusion model (FGDM) that employs frequency-domain filters to guide the diffusion model for structure-preserving image translation. Based on its design, FGDM allows zero-shot learning, as it can be trained solely on the data from the target domain, and used directly for source-to-target domain translation without any exposure to the source-domain data during training. We evaluated it on three cone-beam CT (CBCT)-to-CT translation tasks for different anatomical sites, and a cross-institutional MR imaging translation task. FGDM outperformed the state-of-the-art methods (GAN-based, VAE-based, and diffusion-based) in metrics of Fréchet Inception Distance (FID), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM), showing its significant advantages in zero-shot medical image translation.