<i>Objective.</i>This study investigates key factors influencing deep learning-based dose prediction models for head and neck cancer radiation therapy. The goal is to evaluate model accuracy, robustness, and computational efficiency, and to identify key components necessary for optimal performance.<i>Approach.</i>We systematically analyze the impact of input and dose grid resolution, input type, loss function, model architecture, and noise on model performance. Two datasets are used: a public dataset (OpenKBP) and an in-house clinical dataset. Model performance is primarily evaluated using two metrics: dose score and dose-volume histogram (DVH) score.<i>Main results.</i>High-resolution inputs improve prediction accuracy (dose score and DVH score) by 8.6%-13.5% compared to low resolution. Using a combination of CT, planning target volumes, and organs-at-risk as input significantly enhances accuracy, with improvements of 57.4%-86.8% over using CT alone. Integrating mean absolute error (MAE) loss with value-based and criteria-based DVH loss functions further boosts DVH score by 7.2%-7.5% compared to MAE loss alone. In the robustness analysis, most models show minimal degradation under Poisson noise (0-0.3 Gy) but are more susceptible to adversarial noise (0.2-7.8 Gy). Notably, certain models, such as SwinUNETR, demonstrate superior robustness against adversarial perturbations.<i>Significance.</i>These findings highlight the importance of optimizing deep learning models and provide valuable guidance for achieving more accurate and reliable radiotherapy dose prediction.

Cone-beam CT (CBCT) is crucial for patient alignment and target verification in radiation therapy (RT). However, for non-coplanar beams, potential collisions between the treatment couch and the on-board imaging system limit the range that the gantry can be rotated. Limited-angle measurements are often insufficient to generate high-quality volumetric images for image-domain registration, therefore limiting the use of CBCT for position verification. An alternative to image-domain registration is to use a few 2D projections acquired by the onboard kV imager to register with the 3D planning CT for patient position verification, which is referred to as 2D-3D registration. The 2D-3D registration involves converting the 3D volume into a set of digitally reconstructed radiographs (DRRs) expected to be comparable to the acquired 2D projections. The domain gap between the generated DRRs and the acquired projections can happen due to the inaccurate geometry modeling in DRR generation and artifacts in the actual acquisitions. We aim to improve the efficiency and accuracy of the challenging 2D-3D registration problem in non-coplanar RT with limited-angle CBCT scans. We designed an accelerated, dataset-free, and patient-specific 2D-3D registration framework based on an implicit neural representation (INR) network and a composite similarity measure. The INR network consists of a lightweight three-layer multilayer perception followed by average pooling to calculate rigid motion parameters, which are used to transform the original 3D volume to the moving position. The Radon transform and imaging specifications at the moving position are used to generate DRRs with higher accuracy. We designed a composite similarity measure consisting of pixel-wise intensity difference and gradient differences between the generated DRRs and acquired projections to further reduce the impact of their domain gap on registration accuracy. We evaluated the proposed method on both simulation data and real phantom data acquired from a Varian TrueBeam machine. Comparisons with a conventional non-deep-learning registration approach and ablation studies on the composite similarity measure were conducted to demonstrate the efficacy of the proposed method. In the simulation data experiments, two X-ray projections of a head-and-neck image with <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><msup><mn>45</mn> <mo>∘</mo></msup> <annotation>${45}^\circ$</annotation></semantics> </math> discrepancy were used for the registration. The accuracy of the registration results was evaluated on experiments set up at four different moving positions with ground-truth moving parameters. The proposed method achieved sub-millimeter accuracy in translations and sub-degree accuracy in rotations. In the phantom experiments, a head-and-neck phantom was scanned at three different positions involving couch translations and rotations. We achieved translation errors of <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><mrow><mo><</mo> <mn>2</mn> <mspace></mspace> <mi>mm</mi></mrow> <annotation>$< 2\nobreakspace {\rm mm}$</annotation></semantics> </math> and subdegree accuracy for pitch and roll. Experiments on registration using different numbers of projections with varying angle discrepancies demonstrate the improved accuracy and robustness of the proposed method, compared to both the conventional registration approach and the proposed approach without certain components of the composite similarity measure. We proposed a dataset-free lightweight INR-based registration with a composite similarity measure for the challenging 2D-3D registration problem with limited-angle CBCT scans. Comprehensive evaluations of both simulation data and experimental phantom data demonstrated the efficiency, accuracy, and robustness of the proposed method.

Accurately registering breast MR images from different time points enables the alignment of anatomical structures and tracking of tumor progression, supporting more effective breast cancer detection, diagnosis, and treatment planning. However, the complexity of dense tissue and its highly non-rigid nature pose challenges for conventional registration methods, which primarily focus on aligning general structures while overlooking intricate internal details. To address this, we propose \textbf{GuidedMorph}, a novel two-stage registration framework designed to better align dense tissue. In addition to a single-scale network for global structure alignment, we introduce a framework that utilizes dense tissue information to track breast movement. The learned transformation fields are fused by introducing the Dual Spatial Transformer Network (DSTN), improving overall alignment accuracy. A novel warping method based on the Euclidean distance transform (EDT) is also proposed to accurately warp the registered dense tissue and breast masks, preserving fine structural details during deformation. The framework supports paradigms that require external segmentation models and with image data only. It also operates effectively with the VoxelMorph and TransMorph backbones, offering a versatile solution for breast registration. We validate our method on ISPY2 and internal dataset, demonstrating superior performance in dense tissue, overall breast alignment, and breast structural similarity index measure (SSIM), with notable improvements by over 13.01% in dense tissue Dice, 3.13% in breast Dice, and 1.21% in breast SSIM compared to the best learning-based baseline.

The study aimed to develop machine learning (ML) models for pretherapy prediction of absorbed doses (ADs) in kidneys and tumoral lesions for metastatic castration-resistant prostate cancer (mCRPC) patients undergoing [¹⁷⁷Lu]Lu-PSMA-617 (Lu-PSMA) radioligand therapy (RLT). By leveraging radiomic features (RFs) from [⁶⁸Ga]Ga-PSMA-11 (Ga-PSMA) PET/CT scans and clinical biomarkers (CBs), the approach has the potential to improve patient selection and tailor dosimetry-guided therapy. Twenty patients with mCRPC underwent Ga-PSMA PET/CT scans prior to the administration of an initial 6.8±0.4 GBq dose of the first Lu-PSMA RLT cycle. Post-therapy dosimetry involved sequential scintigraphy imaging at approximately 4, 48, and 72 h, along with a SPECT/CT image at around 48 h, to calculate time-integrated activity (TIA) coefficients. Monte Carlo (MC) simulations, leveraging the Geant4 application for tomographic emission (GATE) toolkit, were employed to derive ADs. The ML models were trained using pretherapy RFs from Ga-PSMA PET/CT and CBs as input, while the ADs in kidneys and lesions (n=130), determined using MC simulations from scintigraphy and SPECT imaging, served as the ground truth. Model performance was assessed through leave-one-out cross-validation (LOOCV), with evaluation metrics including R² and root mean squared error (RMSE). The mean delivered ADs were 0.88 ± 0.34 Gy/GBq for kidneys and 2.36 ± 2.10 Gy/GBq for lesions. Combining CBs with the best RFs produced optimal results: the extra trees regressor (ETR) was the best ML model for predicting kidney ADs, achieving an RMSE of 0.11 Gy/GBq and an R² of 0.87. For lesion ADs, the gradient boosting regressor (GBR) performed best, with an RMSE of 1.04 Gy/GBq and an R² of 0.77. Integrating pretherapy Ga-PSMA PET/CT RFs with CBs shows potential in predicting ADs in RLT. To personalize treatment planning and enhance patient stratification, it is crucial to validate these preliminary findings with a larger sample size and an independent cohort.

Percutaneous screw fixation technique in pelvic trauma surgery is an extremely challenging operation that typically requires a trial-and-error insertion process under the guidance of continuous intraoperative X-ray. This process can be simplified by utilizing surgical navigation systems. Understanding the complexity of the intraosseous pelvis corridor is essential for establishing the optimal screw corridor, which further facilitates preoperative planning and intraoperative application. Traditional screw corridor search algorithms necessitate traversing the entrance and exit areas of the screw and calculating the distance from the corridor axis to the bone surface to ascertain the location of the screw. This process is computationally complex, and manual measurement by the physician is time consuming, labor intensive, and empirically dependent. In this study, we propose an automated planning algorithm for pelvic screw corridors based on deep learning deformable registration technology, which can efficiently and accurately identify the optimal screw corridors. Compared to traditional methods, the innovations of this study include: (1) the introduction of corridor safety range constraints on screw positioning, which enhances search efficiency; (2) the application of deep learning deformable registration to facilitate the automatic annotation of the screw entrance and exit areas, as well as the safety range of the corridor; and (3) the development of a highly efficient algorithm for optimal corridor searching, quickly determining the corridor without traversing the entrance and exit areas and enhancing efficiency via a vector-based diameter calculation method. The whole framework of the algorithm consists of three key components: atlas generation module, deformable registration and optimal corridor searching strategy. In the experiments, we test the performance of the proposed algorithm on 198 intact pelvises for calculating the optimal corridor of anterior column corridor and S1 sacroiliac screws. The results show that the new algorithm can increase the corridor diameter by 2.1%-3.3% compared to manual measurements, while significantly reducing the average time from 1038s and 3398s to 18.9s and 26.7s on anterior column corridor and S1 sacroiliac corridor, respectively, compared to the traditional screw searching algorithm. This demonstrates the advantages of the algorithm in terms of efficiency and accuracy. However, the current method is validated only on intact pelvises; further research is required for pelvic fracture scenarios.

The purpose of this study is to perform image registration and averaging of multiple free-breathing single-shot cardiac images, where the individual images may have a low signal-to-noise ratio (SNR). To address low SNR encountered in single-shot imaging, especially at low field strengths, we propose a fast deep learning (DL)-based image registration method, called Averaging Morph with Edge Detection (AiM-ED). AiM-ED jointly registers multiple noisy source images to a noisy target image and utilizes a noise-robust pre-trained edge detector to define the training loss. We validate AiM-ED using synthetic late gadolinium enhanced (LGE) images from the MR extended cardiac-torso (MRXCAT) phantom and free-breathing single-shot LGE images from healthy subjects (24 slices) and patients (5 slices) under various levels of added noise. Additionally, we demonstrate the clinical feasibility of AiM-ED by applying it to data from patients (6 slices) scanned on a 0.55T scanner. Compared with a traditional energy-minimization-based image registration method and DL-based VoxelMorph, images registered using AiM-ED exhibit higher values of recovery SNR and three perceptual image quality metrics. An ablation study shows the benefit of both jointly processing multiple source images and using an edge map in AiM-ED. For single-shot LGE imaging, AiM-ED outperforms existing image registration methods in terms of image quality. With fast inference, minimal training data requirements, and robust performance at various noise levels, AiM-ED has the potential to benefit single-shot CMR applications.

We introduce a new type of foundational model for parsing human anatomy in medical images that works for different modalities. It supports supervised or unsupervised training and can perform matching, registration, classification, or segmentation with or without user interaction. We achieve this by training a neural network estimator that maps query locations to atlas coordinates via regression. Efficiency is improved by sparsely sampling the input, enabling response times of less than 1 ms without additional accelerator hardware. We demonstrate the utility of the algorithm in both CT and MRI modalities.

Deformable registration is a fundamental task in medical image processing, aiming to achieve precise alignment by establishing nonlinear correspondences between images. Traditional methods offer good adaptability and interpretability but are limited by computational efficiency. Although deep learning approaches have significantly improved registration speed and accuracy, they often lack flexibility and generalizability across different datasets and tasks. In recent years, foundation models have emerged as a promising direction, leveraging large and diverse datasets to learn universal features and transformation patterns for image registration, thus demonstrating strong cross-task transferability. However, these models still face challenges in generalization and robustness when encountering novel anatomical structures, varying imaging conditions, or unseen modalities. To address these limitations, this paper incorporates Sharpness-Aware Minimization (SAM) into foundation models to enhance their generalization and robustness in medical image registration. By optimizing the flatness of the loss landscape, SAM improves model stability across diverse data distributions and strengthens its ability to handle complex clinical scenarios. Experimental results show that foundation models integrated with SAM achieve significant improvements in cross-dataset registration performance, offering new insights for the advancement of medical image registration technology. Our code is available at https://github.com/Promise13/fm_sam}{https://github.com/Promise13/fm\_sam.

Microwave thermotherapy is a promising approach for cancer treatment, but accurate noninvasive temperature monitoring remains challenging. This study aims to achieve accurate temperature prediction during microwave thermotherapy by efficiently integrating multi-feature data, thereby improving the accuracy and reliability of noninvasive thermometry techniques. We proposed an enhanced recurrent neural network architecture, namely CirnetamorNet. The experimental data acquisition system is developed by using the material that simulates the characteristics of human tissue to construct the body model. Ultrasonic image data at different temperatures were collected, and 5 parameters with high temperature correlation were extracted from gray scale covariance matrix and Homodyned-K distribution. Using multi-feature data as input and temperature prediction as output, the CirnetamorNet model is constructed by multi-head attention mechanism. Model performance was evaluated by analyzing training losses, predicting mean square error and accuracy, and ablation experiments were performed to evaluate the contribution of each module. Compared with common models, the CirnetamorNet model performs well, with training losses as low as 1.4589 and mean square error of only 0.1856. Its temperature prediction accuracy of 0.3°C exceeds that of many advanced models. Ablation experiments show that the removal of any key module of the model will lead to performance degradation, which proves that the collaboration of all modules is significant for improving the performance of the model. The proposed CirnetamorNet model exhibits exceptional performance in noninvasive thermometry for microwave thermotherapy. It offers a novel approach to multi-feature data fusion in the medical field and holds significant practical application value.

Reorienting cardiac positron emission tomography (PET) images to the transaxial plane is essential for cardiac PET image analysis. This study aims to design a convolutional neural network (CNN) for automatic reorientation and evaluate its generalizability. An artificial intelligence (AI) method integrating U-Net and the differentiable spatial to numeric transform module (DSNT-U) was proposed to automatically position three feature points (P_apex, P_base, and P_RV), with these three points manually located by an experienced radiologist as the reference standard (RS). A second radiologist performed manual location for reproducibility evaluation. The DSNT-U, initially trained and tested on a [¹¹C]acetate dataset (training/testing: 40/17), was further compared with a CNN-spatial transformer network (CNN-STN). The network fine-tuned with 4 subjects was tested on a [¹³N]ammonia dataset (n = 30). The performance of the DSNT-U was evaluated in terms of coordinates, volume, and quantitative indexes (pharmacokinetic parameters and total perfusion deficit). The proposed DSNT-U successfully achieved automatic myocardial reorientation for both [¹¹C]acetate and [¹³N]ammonia datasets. For the former dataset, the intraclass correlation coefficients (ICCs) between the coordinates predicted by the DSNT-U and the RS exceeded 0.876. The average normalized mean squared error (NMSE) between the short-axis (SA) images obtained through DSNT-U-based reorientation and the reference SA images was 0.051 ± 0.043. For pharmacokinetic parameters, the R² between the DSNT-U and the RS was larger than 0.968. Compared with the CNN-STN, the DSNT-U demonstrated a higher ICC between the estimated rigid transformation parameters and the RS. After fine-tuning on the [¹³N]ammonia dataset, the average NMSE between the SA images reoriented by the DSNT-U and the reference SA images was 0.056 ± 0.046. The ICC between the total perfusion deficit (TPD) values computed from DSNT-U-derived images and the reference values was 0.981. Furthermore, no significant differences were observed in the performance of the DSNT-U prediction among subjects with different genders or varying myocardial perfusion defect (MPD) statuses. The proposed DSNT-U can accurately position P_apex, P_base, and P_RV on the [¹¹C]acetate dataset. After fine-tuning, the positioning model can be applied to the [¹³N]ammonia perfusion dataset, demonstrating good generalization performance. This method can adapt to data of different genders (with or without MPD) and different tracers, displaying the potential to replace manual operations.