Continual Test-Time Adaptation (CTTA) effectively mitigates domain shifts that occur across medical institutions by allowing models to adapt at test time without revisiting the source data. Existing CTTA methods exploit transform invariance through pixel-level augmentation, but these approaches tend to produce inconsistent results in online environments, particularly critical in semantic segmentation tasks. To address these issues, we propose FAWA, a CTTA framework that combines feature-based augmentation and a weighted averaging strategy. FAWA augments features by retrieving statistical information from a continuously updated memory bank, which generates diverse perturbations of input target samples. Each transformed sample is independently fine-tuned, and the resulting parameter sets are merged through continuous weighted averaging. This process leads the model to converge to a stable minimum and mitigates the accumulated error. Experiments on Brain CT and Prostate MRI benchmarks indicate that FAWA achieves competitive performance while ensuring robustness during extended adaptation streams.

Conventional hematoxylin and eosin (H E) staining is destructive and largely limited to two-dimensional sections. We aimed to develop a practical virtual staining method that produces H E-like images from label-free three-dimensional holotomography (HT) while preserving nuclear morphology and requiring only HT at inference. We designed a DAPI-guided conditional diffusion model with a shared encoder and two decoder heads (H E and DAPI). During training, the model receives HT as condition input and predicts diffusion noise for both H E and DAPI targets using mean-squared-error objectives. DAPI is used only during training as nucleus-centric guidance. Data were acquired from the same tissue using a sequential protocol (HT imaging, then DAPI imaging, then H E imaging). Because local nonlinear tissue deformation remains after global affine registration, HT–H E pairs were treated as weakly paired, while HT–DAPI provided stronger local correspondence. Compared with CycleGAN baselines, the proposed model produced more realistic nuclear morphology and better structural consistency. On held-out test tiles, DAPI-guided diffusion achieved lower FID and KID (FID: 9.2158; KID: 0.0091 ± 0.0035 ) than CycleGAN without DAPI (FID: 14.7447; KID: 0.0434 ± 0.0067 ). Training-only DAPI guidance improves virtual H E generation from label-free HT without requiring DAPI during inference. This weakly paired training design reduces dependence on expensive pixel-level registration and supports scalable, nondestructive digital histopathology workflows.

BACKGROUND:Boron Neutron Capture Therapy (BNCT) utilizes high linear energy transfer (LET) charged particles from the 10B(n,α)7Li reaction to selectively destroy tumor cells. Unlike conventional radiation therapy, BNCT facilities rely on fixed beam ports, making patient positioning critical in treatment planning. Current treatment planning, however, depends on manual forward planning using computationally intensive Monte Carlo (MC) simulations. PURPOSE:We propose an automated patient position optimization framework for BNCT to address these limitations. This method overcomes the computational bottleneck of MC simulations by integrating a rapid deep learning (DL) dose prediction model with a high-fidelity GPU-accelerated MC engine, enabling efficient inverse planning. METHODS:We developed a hybrid coarse-to-fine optimization workflow driven by the Trust Region Bayesian Optimization (TuRBO) algorithm. In the coarse stage, a 3D U-Net architecture rapidly identifies promising patient positions. In the fine stage, the solution is refined within a restricted parameter bound using the GPU-accelerated MC engine. We evaluated the framework on the GLIS-RT open dataset (229 glioblastoma patients, split into train, validation, and test sets at a 70:20:10 ratio). RESULTS:Optimizing the patient position with TuRBO improved target dose and homogeneity over a non-optimized baseline, reaching equivalent plan quality about 5.65× faster than an exhaustive grid search. In the coarse stage, the DL dose predictor ran about 37× faster than the high-fidelity MC. The coarse-to-fine workflow then matched high-fidelity MC plan quality while cutting optimization time by a factor of 2.7, outperforming optimization driven by the DL predictor alone. Two-port plans required no change to the method, where the added geometric freedom raised the mean target dose from 59.96% to 74.91% and lowered the homogeneity index (HI) from 1.226 to 0.647. CONCLUSIONS:We successfully developed a Bayesian optimization (BO) framework for BNCT patient positioning. The proposed coarse-to-fine approach effectively balances computational speed with accuracy, offering a practical way toward automated inverse planning in BNCT.

The nonlinear collision operator consumes a significant amount of computation time in tokamak whole-volume modeling, and in current numerical methods, the computational time grows O(n2), with n representing the number of plasma species. In this study, we address the acceleration of the Fokker–Planck–Landau (FPL) collision operator using deep learning techniques. The developed FPL-net, a deep learning-based nonlinear Fokker–Planck–Landau collision operator, is a fully convolutional neural network optimized for computational speed with a compact model structure. FPL-net was trained on data representing various temperature conditions of an electron plasma on a two-dimensional velocity grid, ensuring generality. The network's training incorporated physics-informed loss functions for density, momentum, and energy moments of the plasma probability distribution function, which served as constraints, and it was trained to recursively predict two time steps, achieving robust accuracy. Notably, FPL-net demonstrated full temperature relaxation, representing the first time this has been accomplished by a deep learning–based FPL collision operator. Additional experiments with noisy inputs and extended rollouts validated the model's accuracy, which also shows over 1000x acceleration compared to traditional finite volume methods. We discuss the achieved acceleration through deep learning techniques and propose potential avenues for further enhancement and refinement in future research.

In this study, we investigated the metal artifact reduction (MAR) performance of a deep learning (DL)-based technique in the evaluation of postoperative CT after total knee arthroplasty (TKA). For the development dataset, we collected CT scans from fifty patients without a metal prosthesis, and for the clinical test dataset, we collected CT scans from 44 patients with a previous history of TKA. We developed a DL-based knee MAR network (KMAR-Net) using 25,000 pairs of simulated images generated from 50 patients using the sinogram handling method. Regarding quantitative analysis, the area, mean attenuation, and standard deviation were calculated for Non-MAR, MAR algorithm for orthopedic implants (O-MAR), and KMAR-Net. For qualitative analysis, overall artifact, bone conspicuity, and soft tissue were compared using visual grading analysis. To additionally validate the feasibility of KMAR-Net under controlled conditions, a phantom study using a CTDI phantom with various metallic inserts and scanning parameters was conducted. KMAR-Net outperformed the projection-completion method regarding the area of dark streak artifacts, mean attenuation, and standard deviation within the artifacts. In the qualitative analysis, KMAR-Net was superior to O-MAR in the overall artifact and soft tissue evaluation, and one of the two readers evaluated it as superior for bone conspicuity (P = 0.080 for reader 1 and P < 0.001 for reader 2). In summary, DL-based KMAR-Net showed superior MAR performance in CT compared to the conventional projection-based method.

Despite significant advancements in deep learning for medical screening, generating realistic annotation labels for specific disease groups remains a critical challenge. To address this, we propose an inpainting framework that employs realistic annotation masks, integrating channel-specific conditions across the image volume to produce high-quality paired datasets. Specifically, free-form annotation masks for 2D normal MR images were generated using a variational autoencoder (VAE) adjusted along the z-axis. Additionally, localized areas were inpainted using generative adversarial networks (GANs) with a cascaded generator featuring mask- guided boundary attention and a discriminator operating in a hyperspherical embedding space. This approach ensures the synthesis of high-coverage tumors with seamless integration into adjacent tissues. The proposed framework outperformed existing models, achieving a low FID score of 32.43, demonstrating its ability to generate highly realistic datasets. Moreover, when applied to data augmentation for downstream segmentation tasks, it improved the Dice score from 0.749 to 0.780, demonstrating its potential to enhance segmentation performance. These results underscore the framework's effectiveness in addressing critical challenges in building paired datasets in the medical domain. Our trained model and inference code are available on github.com/kyurikeem/Tumor-Inpainting.

Deep learning-based approaches to metal artifact reduction have recently been proposed, yet these methods still struggle to effectively remove metal artifacts in head and neck computed tomography (CT) images, which have a complex structure and can contain strong artifacts due to the insertion of dental fillings and implants. These strong metal artifacts cause treatment uncertainty in radiation therapy. In this study, we propose a masked criterion function that weighs each region for CT numbers using masks extracted by supervised contrastive learning to better remove metal artifacts in head and neck CT images. Applying the criterion function, a convolutional neural network-based metal artifact reduction model was trained on a synthetic dataset. We adopted a new data synthesis method to prevent tissue information loss by sinogram handling. For the synthetic data, our method outperformed previous models (e.g., linear interpolation, UNet, IndudoNet, FusionNet, Uformer) in terms of image quality and quantitative evaluations, showing the lowest average value of calculated artifact index, 26.57311, compared to the others. In addition, we recalculated the dose on artifact-reduced CT images and found that artifacts clearly degraded the plan quality for patients whose target is close to metal. The results of this study demonstrate that the proposed criterion function helps separate artifacts and tissues using masks extracted through supervised contrastive learning, and that the proposed model can reduce even strong artifacts using this criterion function. Our code can be found here: github.com/wonjin0403/MAR.git

Training deep neural networks with multi-domain data generally gives more robustness and accuracy than training with single domain data, leading to the development of many deep learning-based algorithms using multi-domain data. However, if part of the input data is unavailable due to missing or corrupted data, a significant bias can occur, a problem that may be relatively more critical in medical applications where patients may be negatively affected. In this study, we propose the Laplacian filter attention with style transfer generative adversarial network (LASTGAN) to solve the problem of missing sequences in brain tumor magnetic resonance imaging (MRI). Our method combines image imputation and image-to-image translation to accurately synthesize specific sequences of missing MR images. LASTGAN can accurately synthesize both overall anatomical structures and tumor regions of the brain in MR images by employing a novel attention module that utilizes a Laplacian filter. Additionally, among the other sub-networks, the generator injects a style vector of the missing domain that is subsequently inferred by the style encoder, while the style mapper assists the generator in synthesizing domain-specific images. We show that the proposed model, LASTGAN, synthesizes high quality MR images with respect to other existing GAN-based methods. Furthermore, we validate the use of LASTGAN for data imputation or augmentation through segmentation experiments.

Optical Diffraction Tomography (ODT) is a promising technique for three-dimensional imaging, but practical use demands rigorous robustness testing due to real-world noise factors. Despite the growing importance of machine learning safety, robustness in ODT remains underexplored. We propose the first comprehensive robustness testing protocol for ODT-based classifiers by simulating 16 corruption scenarios to create a corrupted dataset. To enhance robustness and accuracy, we introduce CutPix, a data augmentation strategy that balances shape and texture information through fractal pattern mixing and a cut-and-concatenate approach. Our experiments show that CutPix significantly improves robustness under various corrupted environments compared to existing techniques, particularly against pattern noises. All code and corruption simulation scripts are publicly available at https://github.com/NySunShine/odt-robustness-evaluation.

Objective. This study aims to investigate the feasibility of utilizing generative adversarial networks (GANs) to synthesize high-fidelity computed tomography (CT) images from lower-resolution MR images. The goal is to reduce patient exposure to ionizing radiation while maintaining treatment accuracy and accelerating MR image acquisition. The primary focus is to determine the extent to which low-resolution MR images can be utilized to generate high-quality CT images through a systematic study of spatial resolution-dependent magnetic resonance imaging (MRI)-to-CT image conversion. Approach. Paired MRI-CT images were acquired from healthy control and tumor models, generated by injecting MDA-MB-231 and 4T1 tumor cells into the mammary fat pad of nude and BALB/c mice to ensure model diversification. To explore various MRI resolutions, we downscaled the highest-resolution MR image into three lower resolutions. Using a customized U-Net model, we automated region of interest masking for both MRI and CT modalities with precise alignment, achieved through three-dimensional affine paired MRI-CT registrations. Then our customized models, Nested U-Net GAN and Attention U-Net GAN, were employed to translate low-resolution MR images into high-resolution CT images, followed by evaluation with separate testing datasets. Main Results. Our approach successfully generated high-quality CT images (0.142 mm(2)) from both lower-resolution (0.282 mm(2)) and higher-resolution (0.14(2) mm(2)) MR images, with no statistically significant differences between them, effectively doubling the speed of MR image acquisition. Our customized GANs successfully preserved anatomical details, addressing the typical loss issue seen in other MRI-CT translation techniques across all resolutions of MR image inputs. Significance. This study demonstrates the potential of using low-resolution MR images to generate high-quality CT images, thereby reducing radiation exposure and expediting MRI acquisition while maintaining accuracy for radiotherapy.

The generation of depth images of occlusal dental crowns is complicated by the need for customization in each case. To decrease the workload of skilled dental technicians, various computer vision models have been used to generate realistic occlusal crown depth images with definite crown surface structures that can ultimately be reconstructed to three-dimensional crowns and directly used in patient treatment. However, it has remained difficult to generate images of the structure of dental crowns in a fluid position using computer vision models. In this paper, we propose a two-stage model for generating depth images of occlusal crowns in diverse positions. The model is divided into two parts: segmentation and inpainting to obtain both shape and surface structure accuracy. The segmentation network focuses on the position and size of the crowns, which allows the model to adapt to diverse targets. The inpainting network based on a GAN generates curved structures of the crown surfaces based on the target jaw image and a binary mask made by the segmentation network. The performance of the model is evaluated via quantitative metrics for the area detection and pixel-value metrics. Compared to the baseline model, the proposed method reduced the MSE score from 0.007001 to 0.002618 and increased DICE score from 0.9333 to 0.9648. It indicates that the model showed better performance in terms of the binary mask from the addition of the segmentation network and the internal structure through the use of inpainting networks. Also, the results demonstrated an improved ability of the proposed model to restore realistic details compared to other models.

Generative modeling has seen significant advancements in recent years, especially in the realm of text-to-image synthesis. Despite this progress, the medical field has yet to fully leverage the capabilities of large-scale foundational models for synthetic data generation. This paper introduces a framework for text-conditional magnetic resonance (MR) imaging generation, addressing the complexities associated with multi-modality considerations. The framework comprises a pre-trained large language model, a diffusion-based prompt-conditional image generation architecture, and an additional denoising network for input structural binary masks. Experimental results demonstrate that the proposed framework is capable of generating realistic, high-resolution, and high-fidelity multi-modal MR images that align with medical language text prompts. Further, the study interprets the cross-attention maps of the generated results based on text-conditional statements. The contributions of this research lay a robust foundation for future studies in text-conditional medical image generation and hold significant promise for accelerating advancements in medical imaging research.

MR-only radiotherapy planning is beneficial from the perspective of both time and safety since it uses synthetic CT for radiotherapy dose calculation instead of real CT scans. To elevate the accuracy of treatment planning and apply the results in practice, various methods have been adopted, among which deep learning models for image-to-image translation have shown good performance by retaining domain-invariant structures while changing domain-specific details. In this paper, we present an overview of diverse deep learning approaches to MR-to-CT synthesis, divided into four classes: convolutional neural networks, generative adversarial networks, transformer models, and diffusion models. By comparing each model and analyzing the general approaches applied to this task, the potential of these models and ways to improve the current methods can be can be evaluated.

Echocardiography is the gold standard of diagnosis and evaluation of patent ductus arteriosus (PDA), a common condition among preterm infants that can cause hemodynamic abnormalities and increased mortality rates, but this technique requires a skilled specialist and is not always available. Meanwhile, chest X-ray (CXR) imaging is also known to exhibit signs of PDA and is a routine imaging modality in neonatal intensive care units. In this study, we aim to find and objectively define CXR image features that are associated with PDA by training and visually analyzing a deep learning model. We first collected 4617 echocardiograms from neonatal intensive care unit patients and 17,448 CXR images that were taken 4 days before to 3 days after the echocardiograms were obtained. We trained a deep learning model to predict the presence of severe PDA using the CXR images, and then visualized the model using GradCAM++ to identify the regions of the CXR images important for the model’s prediction. The visualization results showed that the model focused on the regions around the upper thorax, lower left heart, and lower right lung. Based on these results, we hypothesized and evaluated three radiographic features of PDA: cardiothoracic ratio, upper heart width to maximum heart width ratio, and upper heart width to thorax width ratio. We then trained an XGBoost model to predict the presence of severe PDA using these radiographic features combined with clinical features. The model achieved an AUC of 0.74, with a high specificity of 0.94. Our study suggests that the proposed radiographic features of CXR images can be used as an auxiliary tool to predict the presence of PDA in preterm infants. This can be useful for the early detection of PDA in neonatal intensive care units in cases where echocardiography is not available.

This paper presents a novel approach for generating virtual non-contrast planning computed tomography (VNC-pCT) images from contrast-enhanced planning CT (CE-pCT) scans using a deep learning model. Unlike previous studies, which often lacked sufficient data pairs of contrast-enhanced and non-contrast CT images, we trained our model on dual-energy CT (DECT) images, using virtual non-contrast CT (VNC CT) images as outputs instead of true non-contrast CT images. We used a deterministic method to convert CE-pCT images into pseudo DECT images for model application. Model training and evaluation were conducted on 45 patients. The performance of our model, ’VNC-Net’, was evaluated using various metrics, demonstrating high scores for quantitative performance. Moreover, our model accurately replicated target VNC CT images, showing close correspondence in CT numbers. The versatility of our model was further demonstrated by applying it to pseudo VNC DECT generation, followed by conversion to VNC-pCT. CE-pCT images of ten liver cancer patients and ten left-sided breast cancer patients were used. A quantitative comparison with true non-contrast planning CT (TNC-pCT) images validated the accuracy of the generated VNC-pCT images. Furthermore, dose calculations on CE-pCT and VNC-pCT images from patients undergoing volumetric modulated arc therapy for liver and breast cancer treatment showed the clinical relevance of our approach. Despite the model’s overall good performance, limitations remained, particularly in maintaining CT numbers of bone and soft tissue less influenced by contrast agent. Future research should address these challenges to further improve the model’s accuracy and applicability in radiotherapy planning. Overall, our study highlights the potential of deep learning models to improve imaging protocols and accuracy in radiotherapy planning.

During a radiotherapy (RT) course, geometrical variations of target volumes, organs at risk, weight changes (loss/gain), tumor regression and/or progression can significantly affect the treatment outcome. Adaptive RT has become the effective methods along with technical advancements in imaging modalities including cone-beam computed tomography (CBCT). Planning CT (pCT) can be modified via deformable image registration (DIR), which is applied to the pair of pCT and CBCT. However, the artifact existed in both pCT and CBCT is a vulnerable factor in DIR. The dose calculation on CBCT is also suggested. Missing information due to the artifacts hinders the accurate dose calculation on CBCT. In this study, we aim to develop a deep learning-based metal artifact reduction (MAR) model to reduce the metal artifacts in CBCT for head and neck cancer RT. To train the proposed MAR model, we synthesized the kV-CBCT images including metallic implants, with and without metal artifacts (simulated image data pairs) through sinogram image handling process. We propose the deep learning architecture which focuses on both artifact removal and reconstruction of anatomic structure using a dual-encoder architecture. We designed four single-encoder models and three dual-encoder models based on UNet (for an artifact removal) and FusionNet (for a tissue restoration). Each single-encoder model contains either UNet or FusionNet, while the dual-encoder models have both UNet and FusionNet architectures. In the dual-encoder models, we implemented different feature fusion methods, including simple addition, spatial attention, and spatial/channel wise attention. Among the models, a dual-encoder model with spatial/channel wise attention showed the highest scores in terms of peak signal-to-noise ratio, mean squared error, structural similarity index, and Pearson correlation coefficient. CBCT images from 34 head and neck cancer patients were used to test the developed models. The dual-encoder model with spatial/channel wise attention showed the best results in terms of artifact index. By using the proposed model to CBCT, one can achieve more accurate synthetic pCT for head and neck patients as well as better tissue recognition and structure delineation for CBCT image itself.

Objectives We propose a deep learning (DL) multitask learning framework using convolutional neural network for a direct conversion of single-energy CT (SECT) to 3 different parametric maps of dual-energy CT (DECT): virtual-monochromatic image (VMI), effective atomic number (EAN), and relative electron density (RED).Methods We propose VMI-Net for conversion of SECT to 70, 120, and 200 keV VMIs. In addition, EAN-Net and RED-Net were also developed to convert SECT to EAN and RED. We trained and validated our model using 67 patients collected between 2019 and 2020. Single-layer CT images with 120 kVp acquired by the DECT (IQon spectral CT; Philips Healthcare, Amsterdam, Netherlands) were used as input, while the VMIs, EAN, and RED acquired by the same device were used as target. The performance of the DL framework was evaluated by absolute difference (AD) and relative difference (RD).Results The VMI-Net converted 120 kVp SECT to the VMIs with AD of 9.02 Hounsfield Unit, and RD of 0.41% compared to the ground truth VMIs. The ADs of the converted EAN and RED were 0.29 and 0.96, respectively, while the RDs were 1.99% and 0.50% for the converted EAN and RED, respectively.Conclusions SECT images were directly converted to the 3 parametric maps of DECT (ie, VMIs, EAN, and RED). By using this model, one can generate the parametric information from SECT images without DECT device. Our model can help investigate the parametric information from SECT retrospectively.Advances in knowledge DL framework enables converting SECT to various high-quality parametric maps of DECT.

Hematology analysis, a common clinical test for screening various diseases, has conventionally required a chemical staining process that is time-consuming and labor-intensive. To reduce the costs of chemical staining, label-free imaging can be utilized in hematology analysis. In this work, we exploit optical diffraction tomography and the fully convolutional one-stage object detector or FCOS, a deep learning architecture for object detection, to develop a label-free hematology analysis framework. Detected cells are classified into four groups: red blood cell, abnormal red blood cell, platelet, and white blood cell. In the results, the trained object detection model showed superior detection performance for blood cells in refractive index tomograms (0.977 mAP) and also showed high accuracy in the four-class classification of blood cells (0.9708 weighted F1 score, 0.9712 total accuracy). For further verification, mean corpuscular volume (MCV) and mean corpuscular hemoglobin (MCH) were compared with values obtained from reference hematology equipment, with our results showing reasonable correlation in both MCV (0.905) and MCH (0.889). This study provides a successful demonstration of the proposed framework in detecting and classifying blood cells using optical diffraction tomography for label free hematology analysis.

Monitoring tumor volume changes in response to therapeutic agents is a critical step in preclinical drug development. Here, an automated magnetic resonance imaging (MRI)-based approach is proposed using a deep learning framework for tracking longitudinal tumor volume changes in an orthotopic breast cancer model treated with chemotherapy. Longitudinal magnetic resonance images are employed to track changes in tumor volume over time, using an untreated group and a doxorubicin-treated group as the dataset to evaluate treatment effects. Our approach, called Tumor Segmentation-Net (TS-Net), involves replacing the encoder of U-Net with a pre-trained ResNet34 to improve performance. The model was trained using a sample size of n=19 from the untreated group and then subsequently assessed on both the untreated group (n=5) and treated group (n=6). The correlation between the tumor volume determined from the ground truth and that obtained from the trained output was strong ( $\text{R}^{2}$ =0.984, slope=0.996). These results can lead to automated three-dimensional visualization of different longitudinal volume changes with and without treatment. Notably, for small tumors with volumes between 2 and 5 mm 3 , the proposed TS-Net demonstrated an average Dice similarity coefficient score of 0.85, indicating the ability to reliably detect early tumors that may often be missed. Our approach offers a promising tool for preclinical evaluation of tumor volume changes and treatment efficacy in animal models.

Purpose To develop and validate deep radiomics models for the diagnosis of osteoporosis using hip radiographs. Materials and Methods A deep radiomics model was developed using 4924 hip radiographs from 4308 patients (3632 women; mean age, 62 years ± 13 [SD]) obtained between September 2009 and April 2020. Ten deep features, 16 texture features, and three clinical features were used to train the model. T score measured with dual-energy x-ray absorptiometry was used as a reference standard for osteoporosis. Seven deep radiomics models that combined different types of features were developed: clinical (model C); texture (model T); deep (model D); texture and clinical (model TC); deep and clinical (model DC); deep and texture (model DT); and deep, texture, and clinical features (model DTC). A total of 444 hip radiographs obtained between January 2019 and April 2020 from another institution were used for the external test. Six radiologists performed an observer performance test. The area under the receiver operating characteristic curve (AUC) was used to evaluate diagnostic performance. Results For the external test set, model D (AUC, 0.92; 95% CI: 0.89, 0.95) demonstrated higher diagnostic performance than model T (AUC, 0.77; 95% CI: 0.70, 0.83; adjusted P < .001). Model DC (AUC, 0.95; 95% CI: 0.92, 0.97; adjusted P = .03) and model DTC (AUC, 0.95; 95% CI: 0.92, 0.97; adjusted P = .048) showed improved diagnostic performance compared with model D. When observer performance without and with the assistance of the model DTC prediction was compared, performance improved from a mean AUC of 0.77 to 0.87 (P = .002). Conclusion Deep radiomics models using hip radiographs could be used to diagnose osteoporosis with high performance. Keywords: Skeletal-Appendicular, Hip, Absorptiometry/Bone Densitometry © RSNA, 2022