Physics in medicine and biology最新文献

Objective.Online adaptive proton therapy could benefit from reoptimization that considers the total dose delivered in previous fractions. However, the accumulated dose is uncertain because of deformable image registration (DIR) uncertainties. This work aims to evaluate the accuracy of a tool predicting the dose accumulation reliability of a treatment plan, allowing consideration of this reliability during treatment planning.Approach.A previously developed deep-learning-based DIR uncertainty model was extended to calculate theexpectedDIR uncertainty only from the planning computed tomography (CT) and theexpecteddose accumulation uncertainty by including the planned dose distribution. For 5 lung cancer patients, the expected dose accumulation uncertainty was compared to the uncertainty of the accumulated dose of 9 repeated CTs. The model was then applied to several alternative treatment plans for each patient to evaluate its potential for plan selection.Results.The average accumulated dose uncertainty was close to the expected dose uncertainty for a large range of expected uncertainties. For high expected uncertainties, the model slightly overestimated the uncertainty. For individual voxels, errors up to 5% of the prescribed dose were common, mainly due to the daily dose distribution deviating from the plan and not because of inaccuracies in the expected DIR uncertainty. Despite the voxel-wise inaccuracies, the method proved suitable to select and compare treatment plans with respect to their accumulation reliability.Significance.Using our tool to select reliably accumulatable treatment plans can facilitate the use of accumulated doses during online reoptimization.

Objective: We propose a new formulation for ideal observers (IOs) that incorporate stochastic object models (SOMs) for data acquisition optimization. Approach: A data acquisition system is considered as a (possibly nonlinear) discrete-to-discrete mapping from a finite-dimensional object space, x∈R^(n_d), to a finite-dimensional measurement space, y∈R^m. For binary tasks, the two underlying SOMs, H_0 and H_1, are specified by two probability density functions (PDFs) p_0 (x), p_1 (x). This leads to the notion of intrinsic likelihood ratio (LR) Λ_I (x)=p_1 (x)/p_0 (x) and intrinsic class separability (ICS), the latter quantifies the population separability that is independent of data acquisition. With respect to ICS, the IO employs the "extrinsic" LR Λ(y)=pr (y|H_1)/pr(y|H_0) of the data and quantifies the extrinsic class separability (ECS). The difference between ICS and ECS measures the efficiency of data acquisition. We show that the extrinsic LR Λ(y) is the expectation of the intrinsic LR Λ_I (x), where the expectation is with respect to the posterior PDF pr(x│y,H_0 ) under H_0. Main results: We use two examples, one to clarify the new IO and the second to demonstrate its potential for real world applications. Specifically, we apply the new IO to spectral optimization in dual-energy CT projection domain material decomposition (pMD), for which SOMs are used to describe variability of basis material line integrals. The performance rank orders obtained by IO agree with physics predictions. Significance: The main computation in the new IO involves sampling from the posterior PDF pr(x│y,H_0 ), which are similar to (fully) Bayesian reconstruction. Thus our IO computation is amenable to standard techniques already familiar to CT researchers. The example of dual-energy pMD serves as a prototype for other spectral optimization problems, e.g., for photon counting CT or multi-energy CT with multi-layer detectors. .

Objective.To develop an efficient deep learning framework for precise three-dimensional (3D) segmentation of complex orbital structures in multi-sequence magnetic resonance imaging (MRI) and robust assessment of thyroid eye disease (TED) activity, thereby addressing limitations in computational complexity, segmentation accuracy, and integration of multi-sequence features to support clinical decision-making.Approach.We propose RQNet, a U-shaped 3D segmentation network that incorporates the novel Refined Query Transformer Block with refined attention query multi-head self-attention. This design reduces attention complexity fromO(N2)toO(N⋅M)(M≪N) through pooled refined queries. High-quality segmentations then feed into a radiomics pipeline that extracts features per region of interest-including shape, first-order, and texture descriptors. The MRI features from the three sequences-T1-weighted imaging (T1WI), contrast-enhanced T1WI (T1CE), and T2-weighted imaging (T2WI)-are subsequently integrated, with support vector machine, random forest, and logistic regression models employed for assessment to distinguish between active and inactive TED phases.Main results.RQNet achieved Dice similarity coefficients of 83.34%-87.15% on TED datasets (T1WI, T2WI, T1CE), outperforming state-of-the-art models such as nnFormer, UNETR, SwinUNETR, SegResNet, and nnUNet. The radiomics fusion pipeline yielded area under the curve values of 84.65%-85.89% for TED activity assessment, surpassing single-sequence baselines and confirming the benefits of multi-sequence MRI feature fusion enhancements.Significance.The proposed RQNet establishes an efficient segmentation network for 3D orbital MRI, providing accurate depictions of TED structures, robust radiomics-based activity assessment, and enhanced TED assessment through multi-sequence MRI feature integration.

Objective.Radiomics-based artificial intelligence (AI) models show potential in breast cancer diagnosis but lack interpretability. This study bridges the gap between radiomic features (RFs) and Breast Imaging Reporting and Data System (BI-RADS) descriptors through a clinically interpretable framework.Methods. We developed a dual-dictionary approach. First, a clinical mapping dictionary (CMD) was constructed by mapping 56 RFs to BI-RADS descriptors (shape, margin, internal enhancement (IE)) based on literature and expert review. Second, we applied this framework to a classification task to predict triple-negative (TNBC) versus non-TNBC subtypes using dynamic contrast-enhanced MRI data from a multi-institutional cohort of 1549 patients. We trained 27 machine learning classifiers with 27 feature selection methods. Using SHapley Additive exPlanations (SHAP), we interpreted the model's predictions and developed a Statistical Mapping Dictionary for 51 RFs, not included in the CMD.Results. The best-performing model (variance inflation factor feature selector + extra trees classifier) achieved an average cross-validation accuracy of 0.83 ± 0.02. Our dual-dictionary approach successfully translated predictive RFs into understandable clinical concepts. For example, higher values of 'Sphericity', corresponding to a round/oval shape, were predictive of TNBC. Similarly, lower values of 'Busyness', indicating more homogeneous IE, were also associated with TNBC, aligning with existing clinical observations. This framework confirmed known imaging biomarkers and identified novel, data-driven quantitative features.Conclusion.This study introduces a novel dual-dictionary framework (BM1.0) that bridges RFs and the BI-RADS clinical lexicon. By enhancing the interpretability and transparency of AI models, the framework supports greater clinical trust and paves the way for integrating RFs into breast cancer diagnosis and personalized care.

Purpose.Ocular torsion is a challenge occasionally encountered in ocular proton therapy (OPT) consisting of a rotation of the eye about the visual axis. This can result in the safety margin being compromised and reduced conformity of the dose field to the target. This note investigates the effect of ocular torsion on the lateral margin to verify and explore quantitative adaptation strategies to mitigate the adverse effect on this margin.Methods.OCULARIS, an in-house OPT research planning tool, was used to simulate 14 patients undergoing OPT. The lateral margin was determined for each patient at ocular torsion angles ranging from -8^∘to 8^∘in discrete steps of 2^∘, with 19 collimator rotations simulated at each torsion angle.Results.Margin loss increases with greater ocular torsion, with significant inter-patient variability being influenced by the shape of the target. Aligning collimator rotation with ocular torsion nominal torsion matching (NTM) retains 61% of the margin, patient-specific adaptations achieve superior dose conformity to the target. A simple regression method, setting the collimator rotation to the ocular torsion angle minus 1^∘for torsions greater than 2^∘, offers some benefit over NTM in this cohort.Conclusions.Margin loss increases with ocular torsion, with the extent of loss being influenced by patient-specific geometry. The NTM collimator rotation strategy was found to adequately compensate for torsion-induced margin loss. Alternative collimator rotation strategies were also explored, including a framework for optimising collimator rotation in the event of ocular torsion.

Objective.Accurate and personalized radiation dose estimation is crucial for effective targeted radionuclide therapy (TRT). Deep learning (DL) holds promise for this purpose. However, current DL-based dosimetry methods require large-scale supervised data, which is scarce in clinical practice.Approach.To address this challenge, we propose exploring semi-supervised learning (SSL) framework that leverages readily available pre-therapy positron emission tomography (PET) data, where only a small subset requires dose labels, to predict radiation doses, thereby reducing the dependency on extensive labeled datasets. In this study, traditional classification-based SSL approaches were adapted and extended in regression task specifically designed for dose prediction. To facilitate comprehensive testing and validation, we developed a synthetic dataset that simulates PET images and dose calculation using Monte Carlo simulations.Main results.In the experiment, several regression-adapted SSL methods were compared and evaluated under varying proportions of labeled data in the training set. The overall mean absolute percentage error of dose prediction remained between 9% and 11% across different organs, which achieved comparable performance than fully supervised ones.Significance.The preliminary experimental results demonstrated that the proposed SSL methods yield promising outcomes for organ-level dose prediction, particularly in scenarios where clinical data are not available in sufficient quantities.

Objective. Cephalometric alandmark localization is of great clinical significance in diagnosing and treating patients with dental-maxillofacial deformities. Domain shifts across clinical centers significantly hinder model generalizability, causing existing methods to struggle with accurate and robust anatomical landmark localization due to insufficient alignment of high-level semantic features across domains. We aim to improve the cross-domain generalizability of cephalometric landmark detection by aligning semantic features and enhancing output resolution under an unsupervised domain adaptation (UDA) setting.Approach. In this paper, we propose bi-Level alignment with super-resolution head, an effective framework for precise and robust anatomical landmark detection under UDA. Specifically, we employ adaptive instance normalization to generate target-style images while preserving original anatomical spatial structure at the input level. At the output level, a Mean-Teacher framework leverages high-quality pseudo-labels from the teacher model to guide the student model's learning. Additionally, a lightweight super-resolution head enables the generation of high-resolution heatmaps from the multi-scale feature maps and the low-resolution heatmaps, reducing quantization errors with low computational cost.Results. The proposed method achieved a mean localization error of 1.64 mm, a successful detection rate of 72.68% within the clinically acceptable threshold of 2 mm, and an average classification accuracy of 81.81% for anatomical types.Significance. Extensive experiments on public cephalometric datasets demonstrate superiority over state-of-the-art UDA methods, highlighting its potential for clinical applications in cephalometric analysis and orthodontic surgery planning.

Objective.Unsupervised deep learning has shown great promise in deformable image registration (DIR). These methods update model weights to optimize image similarity without requiring ground truth deformation vector fields (DVFs). However, they inherently face the ill-conditioning challenges due to structural ambiguities. This study aims to address these issues by integrating the implicit anatomical understanding of vision foundation models (FMs) into a multi-scale unsupervised framework for accurate and robust DIR.Approach.Our method takes moving and fixed images as inputs and leverages a pre-trained encoder from a vision FM to extract latent features. These features are merged with those extracted by convolutional adaptors to incorporate inductive bias. Correlation-aware multi-layer perceptrons decode the features into DVFs. A pyramid architecture is implemented to capture multi-range dependencies, further enhancing the DIR robustness and accuracy. We evaluated our method using a multi-modality, cross-institutional database consisting of 150 cardiac cine MR and 40 liver CT.Main results.Our model generates realistic and accurate DVFs. Moving images deformed by our method showed excellent similarity to fixed images, achieving a registration Dice score of 0.869 ± 0.093 for cardiac MRI and an average landmark error of 1.60 ± 1.44 mm for liver CT, substantially surpassing the state-of-the-art methods. Ablation studies further verified the effectiveness of integrating foundation features to improve DIR accuracy (p< 0.05).Significance.Our novel approach demonstrates significant advancements in DIR for multi-modality images with complex structures and low contrasts, making it a powerful tool for a wide range of applications in medical image analysis.

Objective.Deformable registration plays a crucial role in motion estimation from a sequence of cardiac magnetic resonance (CMR) imaging, which is good for the diagnosis and treatment of heart diseases. To address the challenges posed by intensity inhomogeneity and complex deformation, we propose a novel convolutional neural network-Transformer framework for this task.Approach.In this study, a convolutional projection Transformer block that enables efficient self-attention computation was designed for modeling long-range spatial correspondences. Additionally, a cooperative learning pattern was adopted for fusing information from global and local features. Finally, multi-resolution strategy was employed for optimizing model parameters in coarse-to-fine manner.Main result.The proposed method was evaluated on three different CMR datasets for intra-subject registration. Experimental results show that the proposed method achieves better Dice overlap and lower surface distance, compared to four non-learning-based methods and three deep-learning-based methods.Significance.For the challenging task of CMR image registration, our method demonstrates superior performance, delivering more accurate results with lower complexity. It may thus facilitate cardiac motion estimation for clinical assessments of cardiac function.

Objective.We investigate scintillation detectors with silicon photomultipliers (SiPMs) as alternatives to direct-conversion detectors based on CdTe/Cd_1-xZn_xTe (CZT) for x-ray photon-counting imaging. Here, we measure counting and spectral performance of three scintillators and compare the results with performances reported in literature for CdTe/CZT detectors for diagnostic photon-counting computed tomography (PCCT).Approach.We built 1 × 1 mm²single-pixel detectors by coupling readily available LYSO:Ce, YAP:Ce, and LaBr₃:Ce scintillators to ultrafast SiPMs. Pulse processing was optimized for rate capability rather than energy resolution. We exposed the detectors to three radioisotopes to determine energy response proportionality and energy resolution. Using an x-ray tube, we measured x-ray spectra and count rate curves, i.e. output count rate (OCR) versus input count rate (ICR).Main results.The energy resolutions of the LYSO:Ce and YAP:Ce detectors exceed 30% full-width-at-half-maximum (FWHM) at 60 keV, with YAP:Ce showing a more proportional response. For a 30 keV count-detection threshold, the maximum OCR of the YAP:Ce detector is 5.4 Mcps pixel^-1for paralyzable-like counting, while the OCR approaches 12.5 Mcps pixel^-1for nonparalyzable-like counting. The LYSO:Ce detector reaches 4.5 Mcps pixel^-1and 10 Mcps pixel^-1, respectively, and the LaBr₃:Ce detector 10.4 Mcps pixel^-1and 22 Mcps pixel^-1. Thereby, the rate capability of the LaBr₃:Ce detector is almost 80% of that reported for two CdTe/CZT detectors for diagnostic PCCT. Moreover, the LaBr₃:Ce detector has high proportionality and an energy resolution of about 20% FWHM at 60 keV, which is comparable to at least one CdTe detector for diagnostic PCCT. The x-ray tube spectra measured using the scintillation detectors show reasonable agreement with incident spectra.Significance.This work indicates that LaBr₃:Ce-based detectors may become an alternative to direct-conversion detectors for diagnostic PCCT, whereas LYSO:Ce- and YAP:Ce-based detectors appear better suited for applications with lower ICR, e.g. cone-beam PCCT in radiotherapy. Ways to further improve x-ray photon-counting scintillation detectors are also discussed.