Robust deformable image registration using synthetic data and transfer learning

The concept of Industry 4.0 uses cyber-physical systems and the Internet of Things to create "smart factories" that enable automated and connected production. However, the complex and diverse nature of cyber-physical systems makes them vulnerable to physical threats, where any fault in their physical components can lead to serious consequences, from financial losses to human safety. To reduce the risk of such failures, it is essential to have techniques in place to monitor the condition of these physical components closely. In recent years, data-driven predictive maintenance methods have become common due to the growth in data and advancements in machine learning. This thesis focuses on fault diagnosis, a key part of predictive maintenance, which aims to detect and identify faults in physical components. Incorporating advanced techniques, including machine learning algorithms and sensor data analysis, enhanced fault diagnosis systems' accuracy and efficiency. However, several real-world barriers challenge the effectiveness of these techniques. This research is structured around three primary objectives. The first objective aims to establish an efficient fault diagnosis system that is highly sensitive to fault occurrences while being resilient to uncertainties such as variability in fault patterns and data imbalance. To address variability in fault patterns, we developed a multi-step approach integrating clustering techniques and an adaptive neuro-fuzzy inference system, which significantly improved diagnosis robustness even under noisy conditions. For imbalanced datasets, a two-step method was introduced, combining permutation entropy for fault detection and envelope analysis for fault isolation. This approach was enhanced by an automatic frequency band selection method leveraging wavelet transform and statistical criteria, resulting in high fault diagnosis accuracy and reliability, even with limited labeled faulty data. The second objective focuses on developing fault diagnosis systems that can maintain high performance despite domain shifts and limited training data in both source and target domains. To address the challenge of domain shifts, a transfer learning technique was introduced, and a systematic literature review was conducted. This study bridges a technical gap by providing a comprehensive taxonomy, state-of-the-art methods, and insights into the challenges of applying transfer learning to predictive maintenance. Additionally, a conceptual framework was proposed to explore scenarios for transferring diagnostic knowledge from different source domains to target domains with insufficient training datasets. In a subsequent study, a new diagnostic framework was developed using digital twins and partial transfer learning to address the lack of real training datasets in both source and target domains as well as label space inconsistencies between them. The framework includes a general approach for developing digital twins of real machines to generate synthetic diagnostic data as source domain data. Furthermore, by incorporating a deep learning network and a double-layer attention mechanism, a partial transfer learning diagnostic model was developed to transfer diagnostic knowledge from the digital twin to the real target domain despite label space inconsistencies. The proposed framework was validated through an industrial case study involving rolling bearings in rotating equipment. The third objective focuses on developing an efficient fault diagnosis system capable of maintaining reliability under unseen and evolving working conditions. To achieve this, a novel self-adaptive fault diagnosis system was proposed, incorporating the MAPE-K loop, a domain generalization network model, and digital twin technologies. This self-adaptive fault diagnosis system leverages the MAPE-K loop to adapt in real time to evolving working conditions. Digital twins were employed to continuously generate data aligned with the real system’s operational context, enabling accurate fault diagnosis in dynamic environments. The domain generalization network model enhances generalization to unseen target working conditions by addressing overfitting and feature loss through multi-domain data augmentation, adversarial learning, and domain discrepancy modules. Validation using multiple rotating machinery datasets demonstrated that the developed fault diagnosis system outperformed state-of-the-art methods across various scenarios. In summary, this dissertation proposes a comprehensive and practical approach to addressing variability in fault patterns, imbalance datasets, data scarcity, domain shifts, and unseen and evolving working conditions challenges in existing fault diagnosis systems.

PurposeKilovoltage cone beam computed tomography (kVCBCT)‐guided adaptive radiation therapy (ART) uses daily deformed CT (dCT), which is generated automatically through deformable registration methods. These registration methods may perform poorly in reproducing volumes of the target organ, rectum, and bladder during treatment. We analyzed the registration errors between the daily kVCBCTs and corresponding dCTs for these organs using the default optical flow algorithm and two registration procedures. We validated the effectiveness of these registration methods in replicating the geometry for dose calculation on kVCBCT for ART.MethodsWe evaluated three deformable image registration (DIR) methods to assess their registration accuracy and dose calculation effeciency in mapping target and critical organs. The DIR methods include (1) default intensity‐based deformable registration, (2) hybrid deformable registration, and (3) a two‐step deformable registration process. Each technique was applied to a computerized imaging reference system (CIRS) phantom (Model 062 M) and to five patients who received volumetric modulated arc therapy to the prostate. Registration accuracy was assessed using the 95% Hausdorff distance (HD95) and Dice similarity coefficient (DSC), and each method was compared with the intensity‐based registration method. The improvement in the dCT image quality of the CIRS phantom and five patients was assessed by comparing dCT with kVCBCT. Image quality quantitative metrics for the phantom included the signal‐to‐noise ratio (SNR), uniformity, and contrast‐to‐noise ratio (CNR), whereas those for the patients included the mean absolute error (MAE), mean error, peak signal‐to‐noise ratio (PSNR), and structural similarity index measure (SSIM). To determine dose metric differences, we used a dose‐volume histogram (DVH) and 3.0%/0.3 mm gamma analysis to compare planning computed tomography (pCT) and kVCBCT recalculations with restimulated CT images used as a reference.ResultsThe dCT images generated by the hybrid (dCTH) and two‐step (dCTC) registration methods resulted in significant improvements compared to kVCBCT in the phantom model. Specifically, the SNR improved by 107% and 107.2%, the uniformity improved by 90% and 75%, and the CNR improved by 212.2% and 225.6 for dCTH and dCTC methods, respectively. For the patient images, the MAEs improved by 98% and 94%, the PSNRs improved by 16.3% and 22.9%, and the SSIMs improved by 1% and 1% in the dCTH and dCTC methods, respectively. For the geometric evaluation, only the two‐step registration method improved registration accuracy. The dCTH method yielded an average HD95 of 12 mm and average DSC of 0.73, whereas dCTC yielded an average HD95 of 2.9 mm and average DSC of 0.902. The DVH showed that the dCTC‐based dose calculations differed by <2% from the expected results for treatment targets and volumes of organs at risk. Additionally, gamma indices for dCTC‐based treatment plans were >95% at all points, whereas they were <95% for kVCBCT‐based treatment plans.ConclusionThe two‐step registration method outperforms the intensity‐based and hybrid registration methods. While the hybrid and two‐step‐based methods improved the image quality of kVCBCT in a linear accelerator, only the two‐step method improved the registration accuracy of the corresponding structures among the pCT and kVCBCT datasets. A two‐step registration process is recommended for applying kVCBCT to ART, which achieves better registration accuracy for local and global image structures. This method appears to be beneficial for radiotherapy dose calculation in patients with pelvic cancer.

Background. To present a new method that determines an optimised IGRT couch correction vector from a displacement vector field (DVF). The DVF is computed by a deformable image registration (DIR) method. The proposed method can improve the quality of volume-of-interest (VOI) alignment in image guided radiation therapy (IGRT), and can serve as a decision-making aid for re-planning. Material and methods. The proposed method was demonstrated using the CT data sets of 11 head-and-neck cancer patients with daily kilovoltage control-CTs. A DVF was computed for each control-CT using a DIR method. The DVF was used for voxel tracking and re-contouring of the VOIs in the control-CTs. Then a rigid body transformation, which could be used as couch correction vector, was optimised. The aim of the optimisation process was to find a vector and rotations that map the deformed VOIs into a specified territory. This territory was defined by a margin extension of the VOIs at the time of the planning process. Within this extension, VOI motion and deformation was tolerated. The objective function in the optimisation process was the sum of all volume fractions outside the defined territories. Results. The proposed method was able to find a correction vector, which resulted in a coverage of the target volumes of at least 98% in 52.3% of all fractions. In contrast, a standard IGRT correction using a rigid registration method only fulfilled this criterion in 22.6% of all fractions. The optimisation process took an average of 1.5 minutes per fraction. Conclusion. The knowledge of the deformation of the anatomy allows the determination of an optimised rigid correction vector using our method. The method ensures controlled mapping of the VOIs despite small deformations. If no optimised vector can be determined, re-planning should be considered. Thus, our method can also serve as a decision-making aid for re-planning.

The aim of this study is to evaluate anatomic lung tumor changes and dosimetric consequences utilizing the deformable daily kilovolt (KV) cone-beam computer tomography (CBCT) image registration. Five patients diagnosed with NSCLC were treated with three-dimensional conformal radiotherapy (3D CRT) and 10 daily KV CBCT image sets were acquired for each patient. Each CBCT image and plan CT were imported into the deformable image registration (DIR) system. The plan CT image was deformed by the DIR system and a new contour on CBCT was obtained by using the auto-contouring function of the DIR. These contours were individually marked as CBCT f1, CBCT f2,..., and CBCT f10, and imported into a treatment planning system (TPS). The daily CBCT plan was individually generated with the same planning criteria based on new contours. These plans were individually marked as CBCTp1, CBCTp2,..., and CBCTp10, followed by generating a dose accumulation plan (DA plan) in original pCT image contour sets by adding all CBCT plans using Varian Eclipse TPS. The maximum, minimum and mean doses to the plan target volume (PTV) in the 5 DA plans were the same with the CT plans. However, the volume of radiation 5, 10, 20, 30, and 50 Gy of the total lungs in DA plans were less than those of the CT plans. The maximum dose of the spinal cord in the DA plans were average 27.96% less than the CT plans. The mean dose for the left, right, and total lungs in the DA plans were reduced by 13.80%, 23.65%, and 12.96%, respectively. The adaptive 3D CRT based on the deformable registration can reduce the dose to the lung and the spinal cord with the same PTV dose coverage. Moreover, it provides a method for further adaptive radiotherapy exploration.

The purpose of the study is to evaluate the accuracy of two deformable image registration algorithms by examining their influence on the dose summation results obtained using 4DCT (four dimensional computed tomography) dose distributions based on '4D' planned and '4D optimal' IMRT (intensity modulated radiation therapy) plans. For ten lung cancer patients, 4D step and shoot IMRT plans were produced. The breathing cycle was divided into ten parts and for each part a set of CT images was acquired. For each patient the treatment plan was copied to the CTs of each phase and subsequently recalculated. Each phase CT was then registered to the average intensity projection (AIP) CT using a deformable image registration (DIR) algorithm and the composite dose distribution was then calculated by summing up the deformed dose distributions from all the phases ('4D' treatment plan). The '4D optimal' treatment plan was created by producing an optimal plan on the CTs of each phase of the respiratory cycle and summing up the deformed dose distributions from all the phases. The results indicate that it is possible to map the dose distributions of different breathing phases in lung using DIR, and that different DIR methods and target characteristics (motion amplitude, size, location) affect the differences between original plan, '4D' and '4D optimal' dose distributions. Although the '4D optimal' plans were designed to achieve 95% target coverage, both of the used DIR methods failed to translate that coverage in some instances. The same variation between these methods was also observed in the '4D' plan comparison. This study shows that it is feasible to perform an acceptably accurate calculation of the composite deformed dose. However, it is important to account for tumor motion and body deformation especially when the tumor volume is small and/or located in the lower lobe of the lung.

Deformable image registration (DIR) is increasingly used for target volume definition in radiotherapy. However, this method is challenging for postoperative breast cancer patients due to the large deformations and non-correspondence caused by tumor resection and clip insertion. In this study, an improved B-splines based DIR method was developed to address this issue for higher registration accuracy. The conventional B-splines based DIR method was improved with the introduction of point metric and masking technique. The point metric minimizes the distance between 2 point sets with known correspondence for regularization of intensity-based B-splines registration. The masking technique reduces the influence of non-corresponding regions in breast computed tomography (CT) images. Two sets of CT images before and after breast surgery were used for image registration. One set was the diagnostic CT image acquired before surgery, and another set was the planning CT image acquired after surgery for breast cancer radiotherapy. A total of 26 sets of CT images from 13 patients were collected retrospectively for the test. The improved DIR method's registration accuracy was evaluated by target registration error (TRE), the Jacobian determinant, and visual assessment. For soft tissue, the difference in the median TRE between the improved DIR method and the conventional DIR method was statistically significant (2.27 vs. 5.88, P<0.05). The Jacobian determinant of the deformation field was positive for all patients. For visual assessment, the improved DIR method with point metric achieved better matching for soft tissue. The improved DIR method's registration accuracy was higher than the conventional DIR method based on the preliminary results. With point metric and masking technique, the influence of large deformations and non-correspondence on registration between pre- and post-operative CT images can be effectively reduced. Therefore, this method provides a feasible way for target volume definition in postoperative breast cancer radiotherapy treatment planning.

Purpose: In this study, we develop an integrated software platform for adaptive radiation therapy (ART) that combines fast and accurate image registration, segmentation, and dose computation/accumulation methods. Methods: The proposed system consists of three key components; 1) deformable image registration (DIR), 2) automatic segmentation, and 3) dose computation/accumulation. The computationally intensive modules including DIR and dose computation have been implemented on a graphics processing unit (GPU). All required patient-specific data including the planning CT (pCT) with contours, daily cone-beam CTs, and treatment plan are automatically queried and retrieved from their own databases. To improve the accuracy of DIR between pCT and CBCTs, we use the double force demons DIR algorithm in combination with iterative CBCT intensity correction by local intensity histogram matching. Segmentation of daily CBCT is then obtained by propagating contours from the pCT. Daily dose delivered to the patient is computed on the registered pCT by a GPU-accelerated superposition/convolution algorithm. Finally, computed daily doses are accumulated to show the total delivered dose to date. Results: Since the accuracy of DIR critically affects the quality of the other processes, we first evaluated our DIR method on eight head-and-neck cancer cases and compared its performance. Normalized mutual-information (NMI) and normalized cross-correlation (NCC) computed as similarity measures, and our method produced overall NMI of 0.663 and NCC of 0.987, outperforming conventional methods by 3.8% and 1.9%, respectively. Experimental results show that our registration method is more consistent and roust than existing algorithms, and also computationally efficient. Computation time at each fraction took around one minute (30–50 seconds for registration and 15–25 seconds for dose computation). Conclusion: We developed an integrated GPU-accelerated software platform that enables accurate and efficient DIR, auto-segmentation, and dose computation, thus supporting an efficient ART workflow. This work was supported by NIH/NCI under grant R42CA137886.

Ventilation imaging using 4D-CT is a convenient and cost effective functional imaging methodology which might be of value in radiotherapy treatment planning to spare functional lung volumes. To calculate ventilation imaging from 4DCT we must use deformable image registration (DIR). This study investigates the dependence of calculated ventilation on DIR methods. DIR of the normal end expiration and inspiration phases of the 4D-CT images was used to correlate the voxels between the two respiratory phases. Three different DIR algorithms, Optical Flow (OF), Diffeomorphic Demons (DD) and Diffeomorphic Morphons (DM), were retrospectively applied to 10 esophagus and 10 lung cancer cases with 4D-CT image sets that encompassed the entire lung volume. The three ventilation extraction methods were used based on either the Jacobian, the change in volume of the voxel derived from the volume change of a voxel (ΔV) or directly calculated from the Hounsfield Unit (HU). The images were compared using the Dice similarity coefficient comparison. Dependence of ventilation images on the DIR was greater for the Δ V and the Jacobian methods than for the HU method. The Dice similarity coefficient for 20% of low ventilation volume for Δ V was 0.33 ± 0.03 between OF and DM, 0.44 ± 0.05 between OF and DD and 0.51 ± 0.04 between DM and DD. The similarity comparisons for Jacobian was 0.32 ± 0.03, 0.44 ± 0.05 and 0.51 ± 0.04 respectively, and for HU 0.53 ± 0.03, 0.56 ± 0.03 and 0.76 ± 0.04 respectively. Ventilation calculation from 4D-CT demonstrates some degree of dependency on the DIR algorithm employed.KeywordsLung FunctionDeformable Image RegistrationHigh ResolutionVentilation Imaging

Purpose: To present a novel methodology for improving the performance of existing deformable image registration (DIR) algorithms by imposing an inverse consistency constraint (ICC), and achieving physiologically more realistic and robust DIR. Method and Materials: The proposed approach combines forward and reverse mappings of a registration algorithm by an inverse consistency constraint based on fixed point theory. Unlikely most ICC‐DIR algorithms, our method does not require explicitly computing inverse transformations, thus reducing both potential numerical errors and required computations. Here we applied the methodology to the Demons algorithm: inverse consistency Demons (iDemons). However the method is generalized to be applied to many current algorithms, such as accelerated demons and energy minimization methods. The validation and comparison of the iDemons and Demons was carried out on 4DCT clinical lungimaging based on: 1) cross‐correlation; 2) inverse consistency errors (ICE); 3) visual verification; 4) robustness. Results: The registration performance of iDemons was superior to Demons. While both methods achieved similar level of correlation coefficient after 200 iterations, 0.9908 for iDemons and 0.9891 for Demons, the average ICE is reduced by 91%: 0.17mm (iDemons) vs. 1.86mm (Demons), resulting in up to 3mm difference in corresponding contours. Since ICE mostly occurred around organ boundaries, minimizing ICE would improve clinical efficacy of DIR based auto‐contouring. Moreover, the iDemons algorithm was shown to be more robustness and less sensitive to parameter selection, which is significant if DIR is to be utilized clinically on routine basis by non‐specialists. Conclusion: The proposed methodology achieves physiologically more realistic DIR. iDemons achieved significant reduction of ICE compared to Demons, which was significant for organ boundaries. Registration results for iDemons were less sensitive to parameter selection, an important consideration for clinical implementation. This novel inverse consistency constraint represents a generalized formalism, which can be applied to other registration algorithms.

In image-guided adaptive radiotherapy, it is important to have the capability to automatically and accurately delineate the rectal wall, which is a major dose-limiting organ in prostate cancer radiotherapy. As image registration is a process to find the spatial correspondence between two images, a major challenge in intensity-based deformable image registration is to deal with the situation where no correspondence exists for some objects between the two images to be registered. One example is the variation of rectal contents due to the presence and absence of bowel gas. The intensity-based deformable image registration methods alone cannot create the correct spatial transformation if there is no correspondence between the source and target images. In this study we implemented an automatic image intensity modification procedure to create artificial gas pockets in the planning computed tomography (CT) images. A diffusion-based deformable image registration algorithm was developed to use an adaptive smoothing algorithm to better handle large organ deformations. The process was tested in 15 prostate cancer cases and 30 daily CT images containing the largest distended rectums. The manually delineated rectums agreed well with the autodelineated rectums when using the image-intensity modification procedure.

Objective. Clinical implementation of deformable image registration (DIR) requires voxel-based spatial accuracy metrics such as manually identified landmarks, which are challenging to implement for highly mobile gastrointestinal (GI) organs. To address this, patient-specific digital twins (DTs) modeling temporally varying motion were created to assess the accuracy of DIR methods.Approach. A total of 21 motion phases simulating digestive GI motion as 4D image sequences were generated from static 3D patient scans using published analytical GI motion models through a multi-step semi-automated pipeline. Eleven datasets, including six T2-weighted FSE MRI (T2w MRI), two T1-weighted 4D golden-angle stack-of-stars, and three contrast-enhanced computed tomography scans were analyzed. The motion amplitudes of the DTs were assessed against real patient stomach motion amplitudes extracted from independent 4D MRI datasets using hierarchical motion reconstruction. The patient-specific DTs were then used to assess six different DIR methods using target registration error, Dice similarity coefficient (DSC), and the 95th percentile Hausdorff distance using summary metrics and voxel-level granular visualizations. Finally, for a subset of T2w MRI scans collected from patients treated with magnetic resonance-guided radiation therapy, dose distributions were warped and accumulated to assess dose warping errors (DWEs), including evaluations of DIR performance in both low- and high-dose regions for patient-specific error estimation.Main results. Our proposed pipeline synthesized patient-specific DTs modeling realistic GI motion, achieving mean and maximum motion amplitudes and a mean log Jacobian determinant within 0.8 mm and 0.01, respectively, similar to published real-patient gastric motion data. It also enables the extraction of detailed quantitative DIR performance metrics and supports rigorous validation of dose mapping accuracy prior to clinical implementation.Significance. The developed pipeline enables rigorously testing DIR tools for dynamic, anatomically complex regions facilitating granular spatial and dosimetric accuracies.

Purpose:CT ventilation imaging (CTVI) is being used to achieve functionalavoidance lung cancer radiation therapy in three clinical trials(NCT02528942, NCT02308709, NCT02843568). To address the need for common CTVIvalidation tools, we have built the Ventilation And Medical Pulmonary ImageRegistration Evaluation (VAMPIRE) Dataset, and present the results of thefirst VAMPIRE Challenge to compare relative ventilation distributionsbetween different CTVI algorithms and other established ventilation imagingmodalities.Methods:The VAMPIRE Dataset includes 50 pairs of 4DCT scans and correspondingclinical or experimental ventilation scans, referred to as referenceventilation images (RefVIs). The dataset includes 25 humans imaged withGalligas 4DPET/CT, 21 humans imaged with DTPA-SPECT and 4 sheep imaged withXenon-CT. For the VAMPIRE Challenge, 16 subjects were allocated to atraining group (with RefVI provided) and 34 subjects were allocated to avalidation group (with RefVI blinded). 7 research groups downloaded theChallenge dataset and uploaded CTVIs based on deformable image registration(DIR) between the 4DCT inhale/exhale phases. Participants used DIR methodsbroadly classified into B-splines, Free-form, Diffeomorphisms orBiomechanical modeling, with CT ventilation metrics based on the DIRevaluation of volume change, Hounsfield Unit change, or various hybridapproaches. All CTVIs were evaluated against the corresponding RefVI usingthe voxel-wise Spearman coefficient rS, and Dicesimilarity coefficients evaluated for low function lung (DSClow)and high function lung (DSChigh).Results:A total of 37 unique combinations of DIR method and CT ventilationmetric were either submitted by participants directly or derived fromparticipant-submitted DIR motion fields using the in-house software, VESPIR.The rS and DSC results reveal a high degree ofinter-algorithm and inter-subject variability among the validation subjects,with algorithm rankings changing by up to 10 positions depending on thechoice of evaluation metric. The algorithm with the highest overallcross-modality correlations used a biomechanical model based DIR with ahybrid ventilation metric, achieving a median (range) of 0.49(0.27–0.73) for rS, 0.52(0.36–0.67) for DSClow and 0.45 (0.28–0.62) forDSChigh. All other algorithms exhibited at least one negativerS value, and/or one DSC value less than0.5.Conclusions:The VAMPIRE Challenge results demonstrate that the cross-modalitycorrelation between CTVIs and the RefVIs vary not only with the choice ofCTVI algorithm, but also with the choice of RefVI modality, imaging subject,and the evaluation metric used to compare relative ventilationdistributions. This variability may arise from the fact that each of thedifferent CTVI algorithms and RefVI modalities provides a distinctphysiologic measurement. Ultimately this variability, coupled with the lackof a ‘gold standard,’ highlight the ongoing importance offurther validation studies before CTVI can be widely translated fromacademic centers to the clinic. It is hoped that the information gleanedfrom the VAMPIRE Challenge can help inform future validation efforts.

Purpose:Multi‐modality deformable image registration (DIR) for head &amp; neck (HN) radiotherapy is difficult, particularly when matching computed tomography (CT) scans with magnetic resonance imaging (MRI) scans. We hypothesized that the ‘shared information’ between images of different modalities was to be found in some form of edge‐based transformation, and that novel edge‐based DIR methods might outperform standard DIR methods.Methods:We propose a novel method that combines gray‐scale edge‐based morphology and mutual information (MI) in two stages. In the first step, we applied a modification of a previously published mathematical morphology method as an efficient gray scale edge estimator, with denoising function. The results were fed into a MI‐based solver (plastimatch). The method was tested on 5 HN patients with pretreatment CT and MR datasets and associated follow‐up weekly MR scans. The followup MRs showed significant regression in tumor and normal structure volumes as compared to the pretreatment MRs. The MR images used in this study were obtained using fast spin echo based T2w images with a 1 mm isotropic resolution and FOV matching the CT scan.Results:In all cases, the novel edge‐based registration method provided better registration quality than MI‐based DIR using the original CT and MRI images. For example, the mismatch in carotid arteries was reduced from 3–5 mm to within 2 mm. The novel edge‐based method with different registration regulation parameters did not show any distorted deformations as compared to the non‐realistic deformations resulting from MI on the original images. Processing time was 1.3 to 2 times shorter (edge vs. non‐edge). In general, we observed quality improvement and significant calculation time reduction with the new method.Conclusion:Transforming images to an ‘edge‐space,’ if designed appropriately, greatly increases the speed and accuracy of DIR.

Purpose. To develop an optimization scheme and semi-physical regularization method for accurate and efficient deformable image registration (DIR) between magnetic resonance imaging (MRI) and computed tomography (CT) for breast cancer radiation therapy planning. Methods and Materials. A sequentially-applied semi-physical model regularization method with bending energy limited diffeomorphism (BELD) was developed to regularize transformation in free-form (B-spline) DIR. Bony structures were first segmented to account for their rigidity during deformation. Feature points were identified and filtered to be used as a penalty term. Rigidity, BELD, and feature point distance penalty terms are applied sequentially during the optimization process based on mutual information metric. Sample MRI and CT images of breast cancer patients were used to demonstrate the registration. Agreements of feature point pairs and contour transfer between the two image modalities were measured to illustrate the accuracy and efficiency for the developed DIR method. Results. The newly developed tool can accurately register MRI to CT with minimum user intervention for situations with moderate to large deformations. For the sample cases tested, the average feature point pair distance was reduced from 6.7 to 3.2 mm by simple B-spline registration and to 1.8 mm by the newly-developed registration method. Compared to simultaneous optimization of multiple penalty terms, the computation time was shortened by 30% and the accuracy was improved by 15%. The bone area warping artifact in deformed MRI disappeared in the proposed method. Considering the resolution of MR images is 1.3 mm, the resulting feature point pair distance is satisfactory. Conclusion. The proposed DIR method of sequentially applying physical model regularization terms during the optimization process can reduce the pitfalls of local extremity, weighting factor adjustment and multi-resolution optimization stages, therefore improving the accuracy and efficiency of MRI to CT registration even with moderate to large deformations commonly seen in radiation treatment planning with multi-modality images.

Medical image segmentation is vital for clinical diagnosis, yet deep learning models face domain shift challenges when test data distributions differ from training data. Single-source domain generalization (DG) has emerged as a solution to these limitations of deep learning models by training models on a single source domain to generalize to unseen target domains. Current single-source DG methods employ domain randomization (e.g., input/feature-space perturbations) to simulate unseen target domains, but face two key limitations: (1) Texture bias in CNN-based architectures due to their local processing nature, leading to overfitting and poor generalization; (2) Restricted augmented style diversity caused by solely source-dependent feature perturbations and input- or feature-only augmentation, resulting in insufficient coverage of target domain distributions and degraded model's generalizationability. Motivated by the Fourier transform's capability to model global dependency and decouple domain invariant features, we propose Random Domain Generalization (RandDG), a frequency-aware domain randomization method for single-source DG in medical image segmentation. This method enhances generalization ability through coordinated input and feature spaces perturbations while maintaining parameter efficiency via lightweight architectural design in the frequencydomain. The proposed RandDG incorporates a novel Global U-Shape Network (GUNet) segmentation architecture for efficient long-range dependency modeling and a Uniform Low Frequency spectrum Transform (ULoFT) filter for feature-space perturbation to enhance domain generalization ability, facilitating the learning of global domain invariant structural information through consistency constraints within the dual-space randomization framework. More specifically, first, the GUNet architecture combines a lightweight encoder with frequency-domain global filtering and forms a U-shape structure through a skip-connected decoder, utilizing Fourier transforms to capture long-range dependencies while minimizing parameters. Second, the ULoFT filter perturbs feature-space styles by Bernoulli mixing statistics of the source low-frequency amplitude spectrum with uniformly sampled values, thereby reducing sole reliance on source statistics and broadening the coverage of potential target domains. Third, a dual-space randomization framework simultaneously applies input-space style augmentation via global intensity nonlinear augmentation (GIN) filters and feature-space amplitude spectrum perturbation via ULoFT filter, regularized by a consistency loss to enforce domain invariant representationlearning. Experiments were conducted on an abdominal CT-MRI dataset for abdominal multi-organ segmentation from CT to T2 spectral presaturation with inversion recovery (T2-SPIR) MRI and a cross-center prostate dataset for prostate segmentation from one center to others. The proposed RandDG method demonstrated superior generalization ability, significantly improving robustness against domain shifts compared to competitive methods. When tested on unseen domains, the proposed method achieved an average DSC of 87.96% and an average HD of 4.82 mm on the abdominal CT-MRI dataset, as well as an average DSC of 75.95% and an average HD of 8.36 mm on the cross-center prostate dataset. Compared to the UNet baseline, the proposed method achieved an improvement of 10.65% in average DSC and a reduction of 5.23 mm in average HD on the abdominal CT-MRI dataset, as well as 19.36% higher average DSC and 3.39 mm lower average HD on the cross-center prostate dataset. Ablation studies confirmed the effectiveness of each component of the proposed RandDGmethod. The proposed RandDG method effectively addresses the critical challenges of texture bias and limited style diversity in single-source DG for medical image segmentation through frequency-aware dual-space randomization framework. This method shows promise for practical deployment in clinical settings, where target domain data may not always beavailable.