Breast cancer is one of the leading causes of cancer-related mortality among women worldwide. Early detection through mammography screening remains critical for improving survival, yet clinical interpretation of mammograms is challenging. Deep learning approaches have been applied to assist radiologists, but CNN-based models often fail to capture global anatomical context, and Transformer-only methods tend to miss fine lesion details. The aim of this study was to develop a unified global–local multi-view hybrid mammography diagnosis model that improves diagnostic accuracy by combining complementary feature representations. We propose a hybrid framework that integrates Swin Transformer modules for global context modeling with Wavelet-CNN blocks for fine-grained lesion feature extraction. Each of the four standard mammography views (Left Craniocaudal, Left Mediolateral Oblique, Right Craniocaudal, and Right Mediolateral Oblique) was processed independently. A hierarchical fusion strategy was then applied: global and local features are first fused within each view and subsequently aggregated across views using an attention-based pooling mechanism. The model was optimized end-to-end with a composite loss function that supervises the global, local, and cross-view fused representations simultaneously. In addition, a CAM-based sparsity regularization was introduced to enhance lesion-focused interpretability, and a weakly supervised learning strategy was adopted to reduce annotation cost by requiring only image-level labels. The method was evaluated on two large public datasets, VinDr-Mammo and CBIS-DDSM, using AUC, accuracy, and F1 score as evaluation metrics. The proposed model achieved AUC scores of 0.816 and 0.803 on VinDr-Mammo and CBIS-DDSM, respectively, outperforming multiple state-of-the-art comparison models. It also demonstrated superior accuracy and F1 scores on VinDr-Mammo. Saliency map visualizations showed that the model consistently highlighted clinically relevant lesion regions. This work demonstrates that combining Transformer-based global context modeling, Wavelet-CNN local feature analysis, and multi-view fusion leads to improved mammography classification performance. The proposed model provides a robust and interpretable framework that has the potential to enhance clinical decision-making and improve early breast cancer detection in large-scale screening programs.

Cone-beam computed tomography (CBCT) is widely used in image-guided radiotherapy, yet traditional reconstruction methods often introduce noise and streak artifacts, limiting the visibility of low-contrast soft tissue. This study presents a deep learning-based reconstruction framework that replaces the conventional filtering step with a ResUNet model, enabling end-to-end CBCT image enhancement. CBCT projections were simulated from high-quality CT volumes and reconstructed without filtering to generate degraded Raw-CBCT images. These were enhanced using a supervised ResUNet network trained to recover anatomical detail and reduce artifacts. Standard FDK reconstruction was used as a reference. Evaluation using similarity metrics and visual inspection confirmed that the ResUNet-based approach provided clearer soft-tissue boundaries, improved overall image quality, and effectively suppressed noise and artifacts. Compared with the traditional method, the ResUNet model delivered more consistent and reliable reconstructions across varying anatomical regions. This study highlights deep learning as a robust alternative to traditional CBCT reconstruction methods.

BACKGROUND:Accurate segmentation of glioma subregions from multi-parametric MRI (MP-MRI) is critical for clinical management but remains challenging due to tumor heterogeneity and ambiguous tissue boundaries. PURPOSE:This study proposes an uncertainty-guided hybrid segmentation framework that integrates spherical projection-based 2D modeling with localized 3D refinement to improve segmentation fidelity. METHODS:The framework was validated on the BraTS 2020 dataset (N = 369). First, a 2D nnU-Net with spherical projection deformation was employed to generate initial slice-wise predictions. Crucially, prediction variance across multiple spherical projections was utilized to quantify voxel-level uncertainty, highlighting regions of low model confidence. A kernel-based sliding window algorithm then spatially localized 3D subvolumes with high cumulative uncertainty. These targeted regions were subsequently fed into a dedicated 3D nnU-Net for volumetric refinement. Finally, the global 2D predictions and local 3D refinements were adaptively fused using weights optimized via Particle Swarm Optimization. The proposed method was implemented to segment the enhancing tumor (ET), tumor core (TC) and whole tumor (WT). Performance was evaluated against standalone 2D and 3D nnU-Net baselines using the Dice Similarity Coefficient (DSC), HD95, sensitivity, and specificity. RESULTS:The proposed method significantly outperformed 2D and 3D baselines across ET, TC, and WT targets. Notably, it achieved a DSC of 0.8124 for ET (vs. 0.7527 for 2D and 0.7530 for 3D), 0.7499 for TC (vs. 0.7002 for 2D and 0.7027 for 3D), 0.9055 for WT (vs. 0.8989 for 2D and 0.9038 for 3D) and demonstrated consistent gains in HD95 and sensitivity. Quantitative metrics and visualizations confirmed improved spatial coherence and boundary preservation in structurally complex regions. CONCLUSION:By utilizing interpretable uncertainty maps as a spatial attention mechanism, this approach dynamically allocates computational resources to anatomically ambiguous regions. The resulting hybrid framework successfully combines 2D efficiency with 3D contextual accuracy, offering a robust solution for automated glioma segmentation.

BACKGROUND:Medical image analysis has witnessed substantial advancements through recent deep learning (DL) algorithms development. Vision Transformers (ViTs) have emerged as a powerful alternative solution by leveraging self-attention to model both local and global interactions. Despite their promise, ViTs are data-intensive and lack inductive biases, limiting their utility in medical imaging. Conversely, radiomics offers domain-specific, interpretable descriptors of image heterogeneity but lacks scalability and integration with deep learning. This study proposes a unified Radiomics-Embedded Vision Transformer (RE-ViT) framework that combines handcrafted radiomic features and data-driven visual embeddings within a ViT architecture. PURPOSE:To develop and evaluate a RE-ViT framework that integrates radiomics and patch-wise ViT embeddings to improve feature representation for medical image classification across heterogeneous datasets. METHODS:Following the classic ViT design, the input image was first resampled into multiple image patches. For each image patch, handcrafted radiomic features, including intensity, texture, and spatial heterogeneity descriptors, were extracted. Simultaneously, standard patch embeddings were obtained via linear projection of pixel intensities. The two embeddings were averaged, normalized, and combined with positional encodings before being tokenized and processed by a ViT encoder. A learnable token aggregates patch-level information for final classification. The model was evaluated on three publicly available datasets, BUSI (lesion malignancy diagnosis through breast ultrasound), ChestXray2017 (lung pneumonitis diagnosis through chest x-ray), and Retinal OCT (retina disease diagnosis through retinal OCT), using 10-fold cross-validation. Performance metrics included accuracy, macro area under the ROC curve (AUC), sensitivity, and specificity. Ablation studies were implemented to assess the contribution of RE-ViT architectural components on these three clinical problems. Comparative analyses were also conducted against CNN (VGG-16, ResNet) and hybrid (TransMed) models. RESULTS:The proposed RE-ViT model demonstrated consistently robust classification performance across all three medical imaging datasets. In BUSI, RE-ViT achieved an accuracy of 0.848 ± 0.027, AUC of 0.950 ± 0.011, sensitivity of 0.796 ± 0.042, and specificity of 0.905 ± 0.020. In ChestXray2017, it yielded an accuracy of 0.950 ± 0.012, AUC of 0.989 ± 0.004, sensitivity of 0.953 ± 0.010, and specificity of 0.975 ± 0.005. In Retinal OCT, RE-ViT achieved an accuracy of 0.938 ± 0.001, AUC of 0.986 ± 0.001, sensitivity of 0.914 ± 0.023, and specificity of 0.969 ± 0.024. In the comparison studies, the RE-ViT matches or outperforms alternatives. Ablation revealed significant performance drops when removing either radiomics or projection-based embeddings. Attention map visualizations demonstrated imaging modality-specific utilization of radiomics and learned features, with improved localization of clinically relevant regions. CONCLUSIONS:The proposed radiomics-embedded vision transformer was developed for multiple image classification tasks. Current results underscore the potential of our approach to advance other transformer-based medical image classification tasks.

PURPOSE:Lattice radiation therapy is an innovative three-dimensional implementation of spatially fractionated radiation therapy. This study aimed to develop a novel lattice-based proton arc therapy (PAT) technique to deliver a highly modulated peak-valley spatial dose distribution and a modulated linear energy transfer (LET) distribution. METHODS AND MATERIALS:We introduced a PAT lattice algorithm that optimizes three key components, including energy-layer selection, spot assignment, and LET optimization (PATLESL). A simulated annealing algorithm was used to select the optimal energy layers. Meanwhile, an adaptive spot assignment strategy was implemented using a spot-sparsity optimization algorithm based on the primal-dual active set with continuation to achieve an improved peak-to-valley dose ratio (PVDR). Subsequently, LET optimization was achieved using the alternating direction method of multipliers with a minimum monitor unit constraint. Eighteen patients were selected to evaluate plan quality and delivery efficiency compared with the lattice technique based on the previously reported spot-scanning proton arc therapy energy sequence optimization algorithm (PATseq). Feasibility and dosimetric accuracy were further validated using clinical phantom measurements. RESULTS:Compared with PATseq, the PATLESL plan exhibited improved dose and LET distributions with higher delivery efficiency. More specifically, PATLESL achieved steeper dose gradients, improving the PVDR from 16.90 ± 18.04 to 41.85 ± 29.69 (p < 0.01), shortening the beam delivery time by 37.43%, and increasing the target spot coverage by 44.12% while reducing the spot number by 93.78%. It also increased the mean LET in the LTV from 2.84 ± 0.38 to 3.59 ± 0.26 keV/μm (p < 0.01), on average. Meanwhile, the phantom measurement confirmed high agreement between the planned and delivered results. CONCLUSIONS:A novel PAT lattice technique using adaptive spot assignment and LET optimization could deliver an enhanced peak-valley spatial dose distribution. Experimental validation on a clinical proton system confirmed deliverability and dosimetric accuracy, which may facilitate future clinical translation of proton lattice radiotherapy.

Background and purpose:Accurate registration of pretreatment Magnetic Resonance Imaging (MRI) to onboard Cone Beam Computed Tomography (CBCT) is critical for liver Stereotactic Body Radiation Therapy (SBRT) but is challenged by poor CBCT soft-tissue contrast and respiratory motion. We developed and validated PhysMorph, a physics-informed deep learning framework designed to provide rapid, anatomically plausible MR-CBCT image registration of the liver. Materials and methods:We developed PhysMorph, a registration framework that incorporated finite element method (FEM) simulations as biomechanical regularization alongside image similarity metrics. The framework was validated on two datasets: (1) simulated data with a known ground-truth deformation derived from longitudinal MR-Linac scans, and (2) clinical MR-CBCT pairs from liver SBRT patients. Performance was assessed using target registration error (TRE), mean surface distance (MSD), and metrics of biomechanical fidelity. Results:On clinical data, PhysMorph achieved a mean TRE of 2.2 ± 1.4 mm and a MSD of 1.60 ± 0.05 mm, significantly outperforming VoxelMorph (4.11 ± 1.53 mm) and SynthMorph (4.41 ± 1.67 mm) while maintaining high biomechanical fidelity. The framework reduced registration time from over 10 min for conventional finite element methods to 103.4 ms, enabling practical real-time application. Conclusions:PhysMorph enables fast, accurate, and physically realistic registration of pretreatment MRI to on-board CBCT for liver SBRT. By integrating MRI's superior soft-tissue visualization while ensuring anatomical plausibility, our approach facilitates precise tumor localization that could enable smaller planning target volumes and more conformal dose distributions, potentially enhancing tumor control while reducing radiation exposure to healthy tissues.

Rotational symmetry is a defining feature of many tomography systems, including computed tomography (CT) and emission computed tomography (ECT), where detectors are arranged in a circular or periodically rotating configuration. This study revisits the image reconstruction process from the perspective of hardware-induced rotational symmetry and introduces a cyclic group equivariance framework for deep learning-based reconstruction. Specifically, we derive a mathematical correspondence that couples cyclic rotations in the projection domain to discrete rotations in the image domain, both arising from the same cyclic group inherent in the hardware design. This insight also reveals the uniformly distributed circular structure of the projection space. Building on this principle, we provide a cyclic rotation equivariant convolution design method to preserve projection domain symmetry and a cyclic group equivariance regularization approach that enforces consistent rotational transformations across the entire network. We further integrate these modules into a domain transform reconstruction framework and validate them using digital brain phantoms, training on discrete models and testing on more complex and realistic fuzzy variants. Results indicate markedly improved generalization and stability, with fewer artifacts and better detail preservation, especially under data distribution deviation. These findings highlight the potential of cyclic group equivariance as a unifying principle for tomographic reconstruction in rotationally symmetric systems, offering a flexible and interpretable solution for scenarios with limited data.

BACKGROUND AND PURPOSE:Accurate prediction of normal brain dosimetric parameters is crucial for the quality control of single-isocenter multi-target (SIMT) stereotactic radiosurgery (SRS) treatment planning. Reliable dose estimation of normal brain tissue is one of the great indicators to evaluate plan quality and is used as a reference in clinics to improve potentially SIMT SRS treatment planning quality consistency. This study aimed to develop a spherical coordinate-defined deep learning model to predict the dose to a normal brain for SIMT SRS treatment planning. METHODS:By encapsulating the human brain within a sphere, 3D volumetric data of planning target volume (PTVs) can be projected onto this geometry as a 2D spherical representation (in azimuthal and polar angles). A novel deep learning model spherical convolutional neural network (SCNN) was developed based on spherical convolution to predict brain dosimetric evaluators from spherical representation. Utilizing 106 SIMT cases, the model was trained to predict brain V50%, V60%, and V66.7%, corresponding to V10Gy and V12Gy, as key dosimetric indicators. The model prediction performance was evaluated using the coefficient of determination (R2), mean absolute error (MAE), and mean absolute percentage error (MAPE). RESULTS:The SCNN accurately predicted normal brain dosimetric values from the modeled spherical PTV representation, with R2 scores of 0.92 ± 0.05/0.94 ± 0.10/0.93 ± 0.09 for V50%/V60%/V66.7%, respectively. MAEs values were 1.94 ± 1.61 cc/1.23 ± 0.98 cc/1.13 ± 0.99 cc, and MAPEs were 19.79 ± 20.36%/20.79 ± 21.07%/21.15 ± 22.24%, respectively. CONCLUSIONS:The deep learning model provides treatment planners with accurate prediction of dose to normal brain, enabling improved consistency in treatment planning quality. This method can be extended to other brain-related analyses as an efficient data dimension reduction method.

This study presents a unified framework, systematically shows a detailed visual representation and emphasizes the understanding via various key factors throughout physical-chemical-biological stages underlying the irradiation process in FLASH radiotherapy (FLASH-RT) from the review of previous published studies. To develop this framework, we pay close attention to the time scale of irradiation process, to incorporated methodologies from current hypotheses documented by in vitro and vivo studies. Concurrently, the framework illustrates the potential differences in tumor and normal cells induced by FLASH-RT. This synthesis of the literature reveals the potential research points in existing hypotheses which lack of consideration on the entire responses and interactions of each stage from initial physical to biological process. Our interpretation provides both a structured overview and a conceptual explanation, facilitating current understanding and further investigation that could be validated in future experimental settings in FLASH-RT research.

In image-guided radiation therapy (IGRT), deformable image registration between computed tomography (CT) and cone beam computed tomography (CBCT) images remain challenging due to the computational cost of iterative algorithms and the data dependence of supervised deep learning methods. Implicit neural representation (INR) provides a promising alternative, but conventional multilayer perceptron (MLP) might struggle to efficiently represent complex, nonlinear deformations. This study introduces a novel INR-based registration framework that models the deformation as a continuous, time-varying velocity field, parameterized by a Kolmogorov–Arnold Network (KAN) constructed using Jacobi polynomials. To our knowledge, this is the first integration of KAN into medical image registration, establishing a new paradigm beyond standard MLP-based INR. For improved efficiency, the KAN estimates low-dimensional principal components of the velocity field, which are reconstructed via inverse principal component analysis and temporally integrated to derive the final deformation. This approach achieves a ~70% improvement in computational efficiency relative to direct velocity field modeling while ensuring smooth and topology-preserving transformations through velocity regularization. Evaluation on a publicly available pelvic CT–CBCT dataset demonstrates up to 6% improvement in registration accuracy over traditional iterative methods and ~3% over MLP-based INR baselines, indicating the potential of the proposed method as an efficient and generalizable alternative for deformable registration.

Background One of the main challenges of utilizing spot-scanning proton arc therapy (SPArc) is treatment delivery efficiency. Previous studies focus on reducing the number of energy layers by ascending switching to shorten the beam delivery time. However, this is not true of all proton therapy systems. The new energy layer switching system was recently upgraded in the University Medical Center Groningen (UMCG), which enables a fast energy layer ascending switching (ELAS).Purpose We introduce a novel adaptive energy switching SPArc optimization algorithm (SPArc-AES) based on the machine-specific delivery characteristics of proton therapy systems.Methods The SPArc-AES optimization algorithm is based on the polynomial increasing feature of energy layer ascending switching. K-Medoids clustering analysis and simulated annealing algorithm were used to optimize the energy delivery sequence. Ten cases were selected to evaluate the plan quality, plan robustness, and the delivery efficiency compared with the previously SPArc energy sequence optimization algorithm, SPArc_seq.Results Without extra constraints in the energy ascending constraints, the SPArc-AES offers a better plan quality and robustness, while the treatment delivery efficiency was significantly improved compared to the SPArc_seq. More specifically, SPArc-AES effectively shortened the energy layer switching time and the beam delivery time by 34.03% and 31.10%, respectively, while offering better target dose conformality and generally lower dose to organs-at-risk.Conclusions Based on the machine-specific delivery characteristics, we introduced a novel adaptive energy switching algorithm for efficient SPArc optimization, which could significantly improve delivery efficiency while enhancing the plan quality by eliminating no longer necessary constraints on the total number of energy layer ascending switching.

Purpose Stereotactic radiosurgery (SRS) is widely used for brain metastases (BM), but the risk of radionecrosis poses a challenge in post-SRS management. Given the lack of noninvasive imaging methods for distinguishing radionecrosis from recurrence, we aimed to design a deep ensemble learning model that integrates patient clinical features and genomic profiles to identify radionecrosis in patients with BM with post-SRS radiographic progression. Methods and Materials We studied 90 BMs from 62 patients with non-small cell lung cancer, with 27 biopsy-confirmed post-SRS local recurrences. Clinical features and molecular features were collected. A deep neural network (DNN) was trained for radionecrosis/recurrence prediction using the 3-month post-SRS T1+c magnetic resonance imaging. Preceding the binary prediction output, latent variables were extracted as 1024 deep features. An ensemble learning model was then developed, comprising 2 submodels that fused deep features with clinical (“D+C”) or genomic (“D+G”) features. We employed our positional encoding method to optimally fuse the low-dimensional clinical/genomic features with the high-dimensional image features. The postfusion feature in each submodel yielded a logit result after traversing fully connected layers. The ensemble's final output was the synthesized result of these 2 submodels’ logits via logistic regression. Model training employed an 8:2 train/test split, and 10 model versions were developed for robustness evaluation. Performance metrics were compared against image-only DNN model and “D+C” and “D+G” submodels. Results The deep ensemble model showed satisfactory performance on the test set, with the area under the receiver operating characteristic curve (ROCAUC) = 0.91 ± 0.04, sensitivity = 0.87 ± 0.16, specificity = 0.86 ± 0.08, and accuracy = 0.87 ± 0.04. This significantly outperformed the image-only DNN result (ROCAUC = 0.71 ± 0.05, sensitivity = 0.66 ± 0.32). Higher average performance was also observed compared to the “D+C” result (ROCAUC = 0.82 ± 0.03, sensitivity = 0.67 ± 0.17) and “D+G” result (ROCAUC = 0.83 ± 0.02, sensitivity = 0.76 ± 0.22). Conclusions The deep ensemble model achieved the best performance among the models evaluated in this study for distinguishing BM radionecrosis from recurrence using 3-month post-SRS T1+c MR images, clinical features, and genomic features. This highlights the potential of artificial intelligence in clinical decision-making for BM management, warranting further investigation into its clinical applications.

Purpose: Radiation pneumonitis (RP) is a serious complication of intensity-modulated radiation therapy (IMRT) for breast cancer patients, underscoring the need for precise and explainable predictive models. This study presents an Explainable Dual-Omics Filtering (EDOF) model that integrates spatially localized dosiomic and radiomic features for voxel-level RP prediction. Methods: A retrospective cohort of 72 breast cancer patients treated with IMRT was analyzed, including 28 who developed RP. The EDOF model consists of two components: (1) dosiomic filtering, which extracts local dose intensity and spatial distribution features from planning dose maps, and (2) radiomic filtering, which captures texture-based features from pre-treatment CT scans. These features are jointly analyzed using the Explainable Boosting Machine (EBM), a transparent machine learning model that enables feature-specific risk evaluation. Model performance was assessed using five-fold cross-validation, reporting area under the curve (AUC), sensitivity, and specificity. Feature importance was quantified by mean absolute scores, and Partial Dependence Plots (PDPs) were used to visualize nonlinear relationships between RP risk and dual-omic features. Results: The EDOF model achieved strong predictive performance (AUC = 0.95 +- 0.01; sensitivity = 0.81 +- 0.05). The most influential features included dosiomic Intensity Mean, dosiomic Intensity Mean Absolute Deviation, and radiomic SRLGLE. PDPs revealed that RP risk increases beyond 5 Gy and rises sharply between 10-30 Gy, consistent with clinical dose thresholds. SRLGLE also captured structural heterogeneity linked to RP in specific lung regions. Conclusion: The EDOF framework enables spatially resolved, explainable RP prediction and may support personalized radiation planning to mitigate pulmonary toxicity.

Intensity-Modulated Radiation Therapy requires the manual adjustment to numerous treatment plan parameters (TPPs) through a trial-and-error process to deliver precise radiation doses to the target while minimizing exposure to surrounding healthy tissues. The goal is to achieve a dose distribution that adheres to a prescribed plan tailored to each patient. Developing an automated approach to optimize patient-specific prescriptions is valuable in scenarios where trade-off selection is uncertain and varies among patients. This study presents a proof-of-concept artificial intelligence (AI) system based on an Adaptive Neuro-Fuzzy Inference System (ANFIS) to guide IMRT planning and achieve optimal, patient-specific prescriptions in aligned with a radiation oncologist's treatment objectives. We developed an in-house ANFIS-AI system utilizing Prescription Dose (PD) constraints to guide the optimization process toward achievable prescriptions. Mimicking human planning behavior, the AI system adjusts TPPs, represented as dose-volume constraints, to meet the prescribed dose goals. This process is informed by a Fuzzy Inference System (FIS) that incorporates prior knowledge from experienced planners, captured through “if-then” rules based on routine planning adjustments. The innovative aspect of our research lies in employing ANFIS's adaptive network to fine-tune the FIS components (membership functions and rule strengths), thereby enhancing the accuracy of the system. Once calibrated, the AI system modifies TPPs for each patient, progressing through acceptable prescription levels, from restrictive to clinically allowable. The system evaluates dosimetric parameters and compares dose distributions, dose-volume histograms, and dosimetric statistics between the conventional FIS and ANFIS. Results demonstrate that ANFIS consistently met dosimetric goals, outperforming FIS with a 0.7% improvement in mean dose conformity for the planning target volume (PTV) and a 28% reduction in mean dose exposure for organs at risk (OARs) in a C-Shape phantom. In a mock prostate phantom, ANFIS reduced the mean dose by 17.4% for the rectum and by 14.1% for the bladder. These findings highlight ANFIS's potential for efficient, accurate IMRT planning and its integration into clinical workflows.

Background: Deep learning has significantly advanced medical image analysis, with Vision Transformers (ViTs) offering a powerful alternative to convolutional models by modeling long-range dependencies through self-attention. However, ViTs are inherently data-intensive and lack domain-specific inductive biases, limiting their applicability in medical imaging. In contrast, radiomics provides interpretable, handcrafted descriptors of tissue heterogeneity but suffers from limited scalability and integration into end-to-end learning frameworks. In this work, we propose the Radiomics-Embedded Vision Transformer (RE-ViT) that combines radiomic features with data-driven visual embeddings within a ViT backbone. Purpose: To develop a hybrid RE-ViT framework that integrates radiomics and patch-wise ViT embeddings through early fusion, enhancing robustness and performance in medical image classification. Methods: Following the standard ViT pipeline, images were divided into patches. For each patch, handcrafted radiomic features were extracted and fused with linearly projected pixel embeddings. The fused representations were normalized, positionally encoded, and passed to the ViT encoder. A learnable [CLS] token aggregated patch-level information for classification. We evaluated RE-ViT on three public datasets (including BUSI, ChestXray2017, and Retinal OCT) using accuracy, macro AUC, sensitivity, and specificity. RE-ViT was benchmarked against CNN-based (VGG-16, ResNet) and hybrid (TransMed) models. Results: RE-ViT achieved state-of-the-art results: on BUSI, AUC=0.950+/-0.011; on ChestXray2017, AUC=0.989+/-0.004; on Retinal OCT, AUC=0.986+/-0.001, which outperforms other comparison models. Conclusions: The RE-ViT framework effectively integrates radiomics with ViT architectures, demonstrating improved performance and generalizability across multimodal medical image classification tasks.

BACKGROUND:Projection interpolation can be used to reduce streak artifacts caused by sparse sampling in cone-beam computed tomography (CBCT) image reconstruction. Conventional analytical interpolation methods create additional blur and artifacts at locations away from the centers of reconstructed CBCT image slices. We hypothesize that the deep learning (DL) interpolation technique may mitigate these limitations. PURPOSE:The purpose of this study is to develop a DL-based technique that interpolates sparsely sampled real patient CBCT projections before reconstruction and post-process reconstructed images for improved image quality and reduced patient imaging dose. METHODS:Real patient CBCT projections are acquired from projection angles that are not exactly evenly spread. The proposed method first linearly interpolates under-sampled projections according to densely sampled angles. Each set of projections is a projection stack of three axes: x, y, and z that represent width, height, and number of projections. The proposed technique re-slices the stack of interpolated projections along y-axis, and each acquired x-z plane slice is processed by a deep residual U-Net (DRU) model to augment the slice's image quality. The resulting slices are reassembled into a stack of densely-sampled projections to be reconstructed into a CBCT volume with the FDK algorithm. A second DRU model further post-processes the reconstructed CBCT volume to improve the image quality. The proposed technique is compared with conventional linear interpolation on sparsely-sampled real patient CBCT projection data (76, 98, and 136 extracted from 680 projections in half-fan geometry). A quantitative analysis is conducted with metrics of peak-signal-to-noise-ratio (PSNR), structural-similarity-index-measure (SSIM), and root-mean-square-error (RMSE). RESULTS:The PSNR, SSIM, and RMSE results of CBCT reconstructed using DRU optimized projections (98-680 projections in half-fan geometry; 49-340 projections in full-fan geometry) are substantially improved compared to those reconstructed using sparsely sampled projections and those reconstructed using linearly interpolated projections. The application of the DRU post-processing technique further improves the image quality. The reconstructed CBCT with combined workflow using 98 sparsely sampled projections for half-fan geometry and 49 projections for full-fan geometry achieved reasonable image and reduced patient imaging dose by 86%. The reconstruction time is about 15 s in addition to a regular FDK reconstruction. CONCLUSION:The combined workflow is the first DL CBCT projection interpolation technique that is demonstrated to work on real patient projection data. Preliminary results demonstrate that the proposed DL interpolation and post-processing techniques performed well in reducing the artifacts of reconstructed CBCT images using under-sampled patient projections with substantially reduced patient imaging dose.

Purpose: Accurate segmentation of glioma subregions in multi-parametric MRI (MP-MRI) is essential for diagnosis and treatment planning but remains challenging due to tumor heterogeneity and ambiguous boundaries. This study proposes an uncertainty-guided hybrid framework integrating spherical projection-based 2D modeling with targeted 3D refinement to enhance segmentation accuracy and interpretability. Methods: Using the BraTS2020 dataset (369 patients, four-modality MP-MRI), three 2D U-Nets were trained to segment enhancing tumor (ET), tumor core (TC), and whole tumor (WT). Voxel-wise uncertainty was quantified via a spherical projection-based 2D nnU-Net, capturing prediction variance across deformed inputs. A 3D sliding window was used to identify high-uncertainty regions, which were refined using a dedicated 3D nnU-Net. Final outputs combined 2D and 3D predictions through a weighted fusion optimized via Particle Swarm Optimization. Results: The proposed method outperformed standalone 2D and 3D baselines, achieving Dice scores of 0.8124 (ET), 0.7499 (TC), and 0.9055 (WT), with consistent gains in sensitivity and visual coherence. Conclusion: This work presents a novel uncertainty-aware segmentation strategy that adaptively integrates 2D and 3D modeling. By focusing refinement on ambiguous regions, it improves both efficiency and accuracy, offering broad applicability to precision neuro-oncology and other high-stakes medical imaging tasks.

Cone-beam CT (CBCT) is commonly used in treatment imaging, but its limited soft tissue contrast presents challenges for liver tumor localization. As a result, indirect localization methods relying on the liver's boundary are commonly utilized, which have limited accuracy for tumor localization. On-board MRI offers superior soft tissue contrast but is limited by the cost. To address this, we devised a method to generate onboard virtual MRI by integrating pretreatment MRI with onboard CBCT, enhancing liver stereotactic body radiation therapy (SBRT) tumor localization accuracy. We employed a finite element method (FEM) for deformable mapping, deforming prior liver MR images onto CBCT geometry to create a virtual MRI. This hybrid virtual-MRI/CBCT (hMRI-CBCT) approach was evaluated in a pilot study involving 48 patients. The hMRI-CBCT demonstrated superb soft-tissue contrast with clear tumor visualization. Registration accuracy of hMRI-CBCT to planning CT significantly surpasses the onboard CBCT to planning CT registration, particularly for tumors not near the liver boundary, with an average error reduction of 1.53 +/- 2 .16 mm. Our study demonstrated that hybrid MRI/CBCT can apparently reduce localization errors in liver SBRT, potentially improving tumor control and reducing toxicities, and opening avenues for further margin reduction and dose escalation.

Most of conventional 2-dimensional (2D) methods verify dose of multiple targets separately one-by-one for Single-isocenter Multiple-target (SIMT) brain plans, which are inefficient and sub-optimal. This study presented a practical method to verify the dose of 2 targets simultaneously for improved efficiency and accuracy. Fifteen Stereotactic Radiation Therapy (SRT) and sixteen Stereotactic Radiosurgery (SRS) plans were used for this study. Each plan has 2 targets coincide with a plane through the plan isocenter. All plans were created in the Eclipse Treatment Planning System (TPS) using a 6 MV flattening filter free photon beam. A 2D detector array, myQA SRS was used for measurements. It has a spatial resolution of 0.4 mm and an active area of 120 × 140 mm2. It can be rotated along the longitudinal axis with a cylindrical phantom with one-degree precision. All plans were delivered with the detector array centered at the plan isocenter and rotated to intersect 2 targets. Six plans with target separations less than 70 mm were crosschecked with the SRS MapCHECK, which has an active area of 77 × 77 mm2. The measured 2D dose distributions were compared with those calculated from the TPS. Gamma-index analysis was performed using 3%/1 mm criteria and a 10% dose threshold. For all 31 SIMT brain plans measured with myQA SRS, the average and standard deviation of the gamma-passing rate was (96.8 ± 2.2)%. For 15 SRT plans and 16 SRS plans, that was (96.6 ± 2.4)% and (97.0 ± 2.1)%, respectively. For 6 plans crosschecked, the average gamma-passing rates were 96.8% vs. 94.8% with myQA SRS and SRS MapCHECK, respectively. A practical method to verify the dose of 2 targets simultaneously was demonstrated. It offers an efficient way for pretreatment verification of SIMT SRT and SRS plans with improved accuracy.