A review of deep learning-based deformable medical image registration

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

ADNI develops standardized protocols for multi-center comparison and has helped over 1,000 scientific publications explore prevention and treatment of AD since 2004.

“ADNI develops standardized protocols for multi-center comparison and helps over 1,000 scientific publications explore the prevention and treatment of AD since 2004”

• cites [26] Jack, 2008, The alzheimer’s disease neuroimaging initiative (adni): Mri methods, J Magnetic Resonance Imaging: Off J Int Soc Magnetic Resonance Med sourcePartly supported

  The Abstract and Discussion (page numbers unavailable) describe ADNI’s development of standardized MRI methods for multisite studies. The Results mention a protocol-selection meeting in 2004, but our copy does not substantiate the claim that ADNI helped over 1,000 publications explore AD prevention and treatment since then. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The ADNI dataset contains MRI longitudinal image scans in AD patients from 63 sites across the US and Canada.

“The ADNI dataset as one example in brain datasets contains MRI longitudinal image scans in AD patients from 63 sites across the US and Canada.”

• cites [26] Jack, 2008, The alzheimer’s disease neuroimaging initiative (adni): Mri methods, J Magnetic Resonance Imaging: Off J Int Soc Magnetic Resonance Med sourcePartly supported

  The Abstract describes longitudinal MRI scans of participants with Alzheimer’s disease at multiple North American sites. It specifies 55 participating sites, however, not the claimed 63, and does not explicitly identify the countries as the US and Canada. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

May combine two claims.

Wu et al. used convolutional-stacked autoencoder (CAE) to extract discriminative features for 3D deformable brain MRI registration, and registration was performed by optimizing NCC of the two features, achieving improved registration accuracy.

“They adopted convolutional-stacked autoencoder (CAE) to extract the discriminative features for 3D deformable brain MRI registration. Then, the registration was performed by optimizing the NCC of the two features and improved registration accuracy was achieved.”

• cites [56] Wu, 2013 sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

PROMISE12 is the dataset made for the MICCAI prostate segmentation challenge, used by 11 academic teams for various segmentation algorithms and showed promising results.

“PROMISE12 as the example in the prostate dataset is the dataset made for MICCAI prostate segmentation challenge, which was used by 11 academic teams for various segmentation algorithms and showed promising results”

• cites [55] Litjens, 2014, Evaluation of prostate segmentation algorithms for mri: The promise12 challenge, Med Image Anal sourcePartly supported

  The Introduction describes the PROMISE12 dataset and its use in a challenge associated with the MICCAI2012 workshop. The Abstract reports 11 participating teams, varied segmentation methods, and good-to-excellent average algorithm performance. However, the teams came from both academic research groups and industry, not exclusively academia. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

More than 375 publications have used COPDGene to explore assessment and identify biomarkers of COPD since 2009.

“More than 375 publications have used COPDGene to explore the assessment and identify the biomarkers of COPD since 2009”

• cites [47] Regan, 2011, Genetic epidemiology of copd (copdgene) study design, COPD: J Chronic Obstructive Pulmonary Dis sourcePartly supported

  The Conclusion (page unavailable) says COPDGene’s CT phenotyping may generate diagnostic imaging biomarkers, and its data will be available for future studies. The paper does not report that more than 375 publications have used COPDGene since 2009; it describes those uses as prospective. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

Fu et al. designed an MR-TRUS registration network for prostate interventions; the supervision deformation field is generated from population-based FE models from point clouds with biomechanical constraints.

“Recently, Fu et al. (75) designed an MR-TRUS registration network for prostate interventions; the supervision deformation field is generated from population-based FE models from point clouds with biomechanical constraints.”

• cites [75] Fu, 2021, Biomechanically constrained non-rigid mr-trus prostate registration using deep learning based 3d point cloud matching, Med Image Anal sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Fan et al. presented a Birnet for brain MRI image registration by utilizing the deformation field estimated from the traditional registration method.

“Fan et al. (74) presented a Birnet for brain MRI image registration by utilizing the deformation field estimated from the traditional registration method.”

• cites [74] Fan, 2019, Birnet: Brain image registration using dual-supervised fully convolutional networks, Med Image Anal sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Registration is of great significance in radiotherapy; it is performed to calculate the offset of the current target from the planned position, and the offset is used to adjust patient position or radiation beam.

“Moreover, registration is also of great significance in radiotherapy (11, 12); it is performed to calculate the offset of the current target from the planned position, and the offset determined from the registration is used to adjust the patient position or the radiation beam.”

• cites [11] Krilavicius, 2016, Predicting respiratory motion for real-time tumour tracking in radiotherapy sourcePartly supported

  The Introduction describes determining a tumour’s current position and compensating for motion by repositioning the beam; the Abstract also mentions controlling the couch or beam. Neither passage describes registration to calculate an offset from the planned position, so the claim’s registration-specific explanation is unsupported. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [12] Brock, 2017, Use of image registration and fusion algorithms and techniques in radiotherapy: Report of the aapm radiation therapy committee task group no. 132, Med Phys sourcePartly supported

  Section 1.B.3 states that registration determines the patient's offset from the planned position and that the resulting transformation is typically used to adjust patient position. Section 6.D.2 also discusses target-focused registration at treatment delivery. The cited passages do not establish that the offset is used to adjust the radiation beam, so that alternative is unsupported. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Yang et al. proposed to register brain MRI images using a U-Net-like structure in one step.

“At first, Yang et al. (73) proposed to register brain MRI images using a U-Net-like structure in one step.”

• cites [73] Yang, 2016, Fast predictive image registration sourcePartly supported

  The Network Structure section (p. 3) describes an encoder–decoder network using patches from fixed and moving images, and the paper tests it on brain images (Experiments, p. 5). But the network predicts LDDMM initial momentum, from which the deformation is computed by integration (Initial Momentum LDDMM Parameterization, p. 2); the copy does not establish that registration is completed in one step. It calls the encoder VGG-style rather than U-Net-like, though the encoder–decoder description supports that resemblance. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Some details of the selected literature were missing in previous reviews.

“On the other hand, some details of the selected literature were missing.”

• cites [19] Boveiri, 2020, Medical image registration using deep neural networks: A comprehensive review, Medical Image Analysis sourceCould not verify

  The Reference Gathering Methodology (p. 3) says the authors found that their initial search overlooked some references, but it does not say that previous reviews omitted details of their selected literature. Searches for previous reviews and missing literature details found no passage in our copy addressing that claim. (LLM, conf=0.72)

  Experimental — accuracy not yet measured
• cites [20] Fu, 2020, Deep learning in medical image registration: A review, Phys Med Biol sourceCould not verify

  The Introduction says that few previous reviews focused specifically on deep learning in medical image registration, and describes this review's aim to provide detailed statistics. It does not say that previous reviews omitted details of their selected literature. Searches for missing details in previous reviews found no passage addressing that point; page numbers were unavailable in our copy. (LLM, conf=0.73)

  Experimental — accuracy not yet measured
• cites [21] Haskins, 2020, Deep learning in medical image registration: A survey, Mach Vision Appl sourceCould not verify

  The Introduction (p. 3) describes how this survey organizes the methods it covers, and the Conclusion (p. 22) summarizes their challenges. Neither passage says that previous reviews omitted details of their selected literature; searches for that comparison found no supporting passage in our copy. (LLM, conf=0.74)

  Experimental — accuracy not yet measured
• cites [22] Xiao, 2021, A review of deep learning-based three-dimensional medical image registration methods, Quantitative Imaging Med Surg sourceCould not verify

  The Review Article section says that few earlier reviews focused on 3D medical image registration, but it does not say they omitted details of their selected literature. Searches for missing or omitted details in previous reviews found no passage addressing that claim. No page number is available in our copy. (LLM, conf=0.72)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Spatial transformer networks (STNs) introduced in 2015 are one significant advancement beneficial in this era.

“The spatial transformer networks (STNs) (80) introduced in 2015 is one significant advancement that is beneficial in this era.”

• cites [80] Jaderberg, 2015, Spatial transformer networks sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

STN can be inserted anywhere in various networks to execute a spatial transform on an input feature map because it is a completely differentiable module.

“STN can be inserted anywhere in various networks to execute a spatial transform on an input feature map because it is a completely differentiable module.”

• cites [80] Jaderberg, 2015, Spatial transformer networks sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Traditional registration algorithms are constructed as continuous optimization problems or discrete optimization problems.

“The traditional registration algorithms are constructed as continuous optimization problems (15, 16) or discrete optimization problems (17).”

• cites [15] Thirion, 1998, Image matching as a diffusion process: an analogy with maxwell’s demons, Med Image Anal sourceCould not verify

  The Abstract (no page given) describes diffusion-based image matching and relates it to traditional attraction-based methods, but does not classify registration algorithms as continuous or discrete optimization problems. Searches for both optimization terms returned no passage in our copy that addresses that distinction. (LLM, conf=0.70)

  Experimental — accuracy not yet measured
• cites [16] Rueckert, 1999, Nonrigid registration using free-form deformations: application to breast mr images, IEEE Trans Med Imaging sourceNot checked — no full text of the source
• cites [17] Heinrich, 2013, Towards realtime multimodal fusion for image-guided interventions using self-similarities sourceSupported

May combine two claims.

MRI provides better contrast for soft tissues, and CT has clear details of bone.

“magnetic resonance imaging (MRI) provides better contrast for soft tissues, and computed tomography (CT) has clear details of bone (3)”

• cites [3] Huang, 2020, A review of multimodal medical image fusion techniques, Comput Math Methods Med sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

COPDGene contains chest CT scans of more than 10,000 individuals for chronic obstructive pulmonary disease (COPD).

“COPDGene as one example in lung datasets contains chest CT scans of more than 10,000 individuals for chronic obstructive pulmonary disease (COPD).”

• cites [47] Regan, 2011, Genetic epidemiology of copd (copdgene) study design, COPD: J Chronic Obstructive Pulmonary Dis sourcePartly supported

  The Abstract’s Methods/Results section says COPDGene planned to enroll 10,000 smokers, with and without COPD, and obtain chest CT scans from all participants. It does not establish that the dataset already contained scans from more than 10,000 individuals. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

May combine two claims.

Sokooti et al. presented a multi-scale network to learn a deformation field of intra-subject 3D chest CT registration, using random DVF as supervision.

“Sokooti et al. (77) presented a multi-scale network to learn a deformation field of intra-subject 3D chest CT registration. They used random DVF as supervision.”

• cites [77] Sokooti, 2017, Nonrigid image registration using multi-scale 3d convolutional neural network sourceSupported

May combine two claims.

Traditional registration algorithms' computational performance is hampered by high dimension and non-convex properties, and their capability to capture complex deformations is limited.

“Their computational performance, however, is hampered by the high dimension and non-convex properties, and their capability to capture complex deformations is limited (18).”

• cites [18] Schnabel, 2016, Advances and challenges in deformable image registration: From image fusion to complex motion modelling, Medical Image Analysis sourcePartly supported

  Section 2.2 says traditional B-spline registration has difficulty with motion discontinuities and describes newer methods that avoid local minima and reduce computational complexity. The paper also says continuous optimization can be prone to local minima. These passages support limitations in modelling complex motion and aspects of computational difficulty, but they do not attribute poor computational performance specifically to high dimensionality or non-convexity. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Neural ODE models have the advantage of adaptive computing, making them potentially appropriate for application in medical applications.

“Neural ODE models have the advantage of adaptive computing, making them potentially appropriate for application in medical applications.”

• cites [163] Chen, 2018, Neural ordinary differential equations sourceSupported

May combine two claims.

Accessing medical data is not allowed for most researchers since clinical data contain private information and only researchers with authorization could utilize the dataset.

“Finally, accessing medical data is not allowed for most researchers since the clinical data contain private information and only researchers with authorization could utilize the dataset (25).”

• cites [25] Li, 2021, A systematic collection of medical image datasets for deep learning, arXiv sourcePartly supported

  Section 6.1 says most non-medical researchers lack authorization to access clinical data and explains that privacy and ethical restrictions limit access. The citing sentence broadens this to most researchers generally, which the passage does not establish. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

DSC is the most popular evaluation metric, accounting for 46% of all studies.

“Based on Figure 1C , DSC is the most popular metric and accounts for 46% of all the studies.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The review summarized 30 public datasets from deep learning registration studies in Table 1.

“Our review summarized 30 public datasets from deep learning registration studies in Table 1 .”

May combine two claims.

The three most popular ROIs investigated are brain, lung, and prostate, accounting for 47%, 19%, and 14% of all studies.

“The three most popular ROIs being investigated are brain, lung, and the prostate, and each accounts for 47%, 19%, and 14% of all the studies.”

May combine two claims.

MR-MR and CT-CT are the most commonly studied modalities in the selected studies, accounting for 50% and 30% respectively.

“As shown in Figure 1A , MR-MR and CT-CT are the most commonly studied modalities, and each accounts for 50% and 30% of the selected studies”

May combine two claims.

Unsupervised and supervised methods are the two most popular method categories, accounting for 31% and 22% of all studies respectively.

“Unsupervised and supervised methods are two most popular methods and each accounts for 31% and 22% of all the studies.”

May combine two claims.

Most of the summarized datasets consist of brain MR and lung CT images, with other modalities in ROIs such as abdomen and knee also included.

“Most of the datasets consist of brain MR and lung CT images, and other modalities in different ROIs such as the abdomen and knee were also included.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The number of literatures for the deep similarity metric category has decreased, and this trend is projected to continue.

“Nowadays, the number of literatures for this category has decreased, and this trend is projected to continue.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

When training a 3D lung CT registration network, a GPU card with a memory larger than 12 G is necessary.

“when training 3D lung CT registration network, a GPU card with a memory larger than 12 G is necessary.”

May combine two claims.

GAN has the smallest model parameter and the parameter of Transformer expands significantly.

“We can see that GAN has the smallest model parameter and the parameter of Transformer expands significantly.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

It is difficult to interpret learned deep similarity metrics, and the iterative process still limits the use of these methods.

“Moreover, it is difficult to interpret the learned deep similarity metrics, and the iterative process still limits the use of these methods.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Fully supervised methods are widely studied and have notable results.

“The fully supervised methods are widely studied and have notable results.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The formulation in the loss function of the learning network constrains the output DVF to be accurate.

“The formulation in the loss function of the learning network constrains the output DVF to be accurate.”

May combine two claims.

Similarity measures in RL-based registration methods are routinely provided in a traditional way, e.g., MSE, NMI, or LCC, and thus they have limited applications in multimodality registration.

“Different from deep similarity metrics-based methods, the similarity measures in these kinds of methods are routinely provided in a traditional way, e.g., MSE, NMI, or LCC. Thus, they have limited applications in multimodality registration.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The generation of real or synthetic deformation fields is hard, and these deformation fields are different from the real ground truth, which will confine the accuracy and efficiency of these kinds of methods.

“However, the generation of real or synthetic deformation fields is hard, and these deformation fields are different from the real ground truth, which will confine the accuracy and efficiency of these kinds of methods.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Generally, the unsupervised learning framework consists of similarity-based methods and GAN-based methods.

“Generally, the unsupervised learning framework consists of the similarity-based methods and GAN-based methods.”

May combine two claims.

Preparation of the required ground truth is inconvenient and leads to the supervised framework having limitations in generalizing results in different domains and applying various registration tasks.

“Even though different data augmentation techniques and data collection methods have been utilized in supervised learning, the preparation of the required ground truth is inconvenient and leads to the fact that the supervised framework has limitations in generalizing results in different domains and applying various registration tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing).

OAI contains MRI and x-ray longitudinal measurements from 4,796 subjects.

“OAI as the example in the knee dataset contains MRI and x-ray longitudinal measurements from 4,796 subjects”

• cites [48] Eckstein, 2014, Imaging research results from the osteoarthritis initiative (oai): A review and lessons learned 10 years after start of enrolment, Ann rheumatic Dis sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

More than 400 research manuscripts have used OAI to explore assessments and interventions of knee osteoarthritis.

“more than 400 research manuscripts have used the OAI to explore the assessments and interventions of knee osteoarthritis”

• cites [48] Eckstein, 2014, Imaging research results from the osteoarthritis initiative (oai): A review and lessons learned 10 years after start of enrolment, Ann rheumatic Dis sourceNot checked — no full text of the source

May combine two claims.

RL-based methods mainly focus on rigid registration, as it is hard for agents to tackle the large state space generated from deformable vector field.

“Moreover, RL-based methods mainly focus on rigid registration, as it is hard for agents to tackle the large state space generated from deformable vector field.”

Classified as probably not a substantial claim. May combine two claims.

Deep learning-based deformable registration methods have greatly attracted researchers' attention because data-driven methods benefit significantly from a large number of paired or unpaired images compared with traditional methods.

“Recently, deep learning-based deformable registration methods have greatly attracted researchers’ attention, as data-driven methods benefit significantly from a large number of paired/unpaired images when compared with traditional methods.”

• AssumptionRests on an assumption: Data-driven methods benefit significantly from a large number of paired or unpaired images when compared with traditional methods.

Probably not a substantial claim: may not pass S3 (load-bearing).

Various traditional registration methods and toolboxes have been devised, including Elastix and ANTs.

“Various traditional registration methods and toolboxes have been devised over the last few decades, e.g., Elastix (13) and ANTs (14).”

• cites [13] Klein, 2009, Elastix: a toolbox for intensity-based medical image registration, IEEE Trans Med Imaging sourceNot checked — no full text of the source
• cites [14] Avants, 2009, Advanced normalization tools (ants), Insight J sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Real deformation fields are generated from traditional registration, and synthetic deformation fields are obtained from statistical models or random transformations.

“The real deformation fields are generated from traditional registration, and the synthetic deformation fields are obtained from statistical models or random transformations.”

May combine two claims.

Systematic sextant TRUS biopsies have poor imaging quality and lack sharp contrast between cancer and normal tissue, resulting in false-negative rates of up to 30%.

“but for systematic sextant biopsies, its poor imaging quality and lack of sharp contrast between cancer and normal tissue results in false-negative rates of up to 30% (10).”

• cites [10] Norberg, 1997, The sextant protocol for ultrasound-guided core biopsies of the prostate underestimates the presence of cancer, Urology sourceNot checked — no full text of the source

May combine two claims.

Uzunova et al. designed a network for registration between 2D brain MRI and 2D cardiac MRI, with ground truth generated using statistical appearance models (SAMs), adapted FlowNet architecture, and obtained outperforming results.

“Uzunova et al. (78) designed a network for the registration between 2D brain MRI and 2D cardiac MRI. Their ground truth is generated utilizing statistical appearance models (SAMs). They adapted FlowNet (79) architecture and obtained outperforming results.”

• cites [78] Uzunova, 2017, Training cnns for image registration from few samples with model-based data augmentation sourceNot checked — no full text of the source
• cites [79] Dosovitskiy, 2015, Flownet: Learning optical flow with convolutional networks sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

So et al. proposed a novel learning-based metric using Bhattacharyya Distances, where dissimilarity of testing image pairs is calculated by incorporating expected intensity distributions learned from registered training image pairs.

“So et al. (57) proposed a novel learning-based metric by using Bhattacharyya Distances. The dissimilarity of the testing image pairs is calculated by incorporating the expected intensity distributions learned from the registered training image pairs.”

• cites [57] So, 2017, A novel learning-based dissimilarity metric for rigid and non-rigid medical image registration by using bhattacharyya distances, Pattern Recognition sourceNot checked — no full text of the source

May combine two claims.

Rohe et al. adopted a U-Net-like network to predict the deformation field for 3D cardiac MRI volume registration, with supervisions generated from mesh segmentations and SSD as loss; their results are comparable with traditional registration.

“Rohe et al. (76) adopted a U-Net-like network to predict the deformation field for 3D cardiac MRI volume registration. The supervisions are transformations generated from mesh segmentations, and SSD between the supervision and prediction is set as the loss function. Their results are comparable with those from traditional registration.”

• cites [76] Rohé, 2017, Svf-net: learning deformable image registration using shape matching sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

A sufficient number of explicitly annotated ground truths are required for deep similarity metric training, which hinders the development of such approaches.

“However, a sufficient number of explicitly annotated ground truths are required for deep similarity metric training, which hinders the development of such approaches.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

One disadvantage of deep iterative methods is that the registration process is time-consuming and iterative.

“One disadvantage of deep iterative methods is that the registration process is time-consuming and iterative.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Training and assessment of DL models on multiple public datasets could better demonstrate generalizability and clearly compare results on fair benchmarks, including computational time and registration accuracy.

“the training and assessment of DL models on multiple public datasets could also better demonstrate the generalizability of the proposed methods and clearly compare the results on fair benchmarks, including computational time and registration accuracy.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Weakly supervised methods naturally improve the performance of registration by introducing anatomical constraints.

“Weakly supervised methods naturally improve the performance of registration by introducing the anatomical constraints.”

• AssumptionRests on an assumption: Anatomical labels such as segmentations and key points are feasible in practice and credible for training.

Probably not a substantial claim: may not pass S3 (load-bearing).

The constraints are usually performed by penalizing the small and large values of the Jacobian determinant.

“Usually, the constraints are performed by penalizing the small and large values of the Jacobian determinant.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

These benefits encourage researchers to study how to use neural ODEs to optimize the registration of medical images.

“These benefits encourage researchers to study how to use neural ODEs to optimize the registration of medical images.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Transformer is unavailable when the data size is large and GPU memory is limited.

“the Transformer is unavailable when the data size is large and GPU memory is limited.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Unsupervised methods and weakly supervised methods are less reliant on ground truth information and have become hot topics for deep learning-based registration algorithms.

“In contrast, unsupervised methods and weakly supervised methods are less reliant on ground truth information and become hot topics for deep learning-based registration algorithms.”

• AssumptionRests on an assumption: Supervised methods need ground truth supervision that is impractical to obtain.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The evolution of learning-based image registration algorithms has followed a similar path to that of deep learning models.

“The evolution of learning-based image registration algorithms has followed a similar path to that of the deep learning models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Recently, popular networks such as Transformer and contrastive learning are explored for deformable medical image registration and achieved promising results.

“Recently, popular networks, such as Transformer and contrastive learning, are explored for deformable medical image registration and achieved promising results.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Some datasets are still collecting data for further imaging analysis support, such as BraTS since 2012.

“Some datasets are still collecting data for further imaging analysis support, such as the data collected for BraTs since 2012 (37).”

• cites [37] Bakas, 2018, Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge, arXiv sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Neural ODE and diffusion models have been explored for deformable image registration.

“Nowadays, the popular neural ODE and diffusion models have also been explored for deformable image registration.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Analyzing transformations in terms of spatially corresponding patches is challenging in multimodal registration and can negatively impact registration results.

“analyzing transformations in terms of spatially corresponding patches could be challenging in multimodal registration and can negatively impact registration results.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Prevalent and testified public databases are normally chosen by research teams in deep learning-based medical image analysis.

“prevalent and testified public databases are thus normally chosen by research teams in deep learning-based medical image analysis.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

CNNs, SAEs, GANs, DRLs, and deep RNNs make up the majority of the recently developed deep learning algorithms used for medical image registration.

“CNNs, SAEs, GANs, DRLs, and deep RNNs make up the majority of the recently developed deep learning algorithms used for medical image registration”

Probably not a substantial claim: may not pass S3 (load-bearing).

Segmentations are useful in evaluating registration algorithms.

“segmentations are also useful in evaluating the registration algorithms.”

May combine two claims.

Multi-stage coarse-to-fine methods consume large GPU memory and training is time-consuming as the networks are trained separately.

“This kind of method consumes large GPU memory and the training is time-consuming as the networks are trained separately.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Another method for learning large deformation is to construct cascaded networks, in which each cascaded network learns an intermediate deformation field and the source image is recursively and progressively warped by the field, finally aligned to the target image.

“Another method for learning large deformation is to construct cascaded networks, in which each cascaded network learns an intermediate deformation field, and the source image is recursively and progressively warped by the field, finally aligned to the target image.”

The quote only partly states this claim.

In this situation, unsupervised methods are promising to tackle the problems of fully supervised methods.

“In this situation, unsupervised methods are promising to tackle the problems.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The other method to guarantee diffeomorphism is to introduce diffeomorphic integration layers.

“The other method is to introduce diffeomorphic integration layers.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The integration is performed by scaling and squaring the stationary velocity field.

“The integration is performed by scaling and squaring the stationary velocity field.”

Probably not a substantial claim: may not pass S4 (based).

GAN can ensure that the generated registered image has the same intensity distribution as the source image while being similar to the target image in terms of structures.

“GAN can ensure that the generated registered image has the same characteristic (intensity distribution) as the source image while being similar to the target image in terms of structures.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Unsupervised similarity-based methods have become more popular in recent years and are appropriate in many deformable registration tasks as they do not require extra information.

“the unsupervised similarity-based methods have become more popular in recent years, and they are appropriate in many deformable registration tasks as they do not require extra information.”

May combine two claims.

Cascaded networks are still memory-consuming and difficult to converge as all cascaded networks are trained simultaneously.

“This kind of network is still memory-consuming and is difficult to converge as all cascaded networks are trained simultaneously.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

An MRI session generally lasts 2 h under supervision and manipulation of professional researchers, and subjects are almost fully constrained in the specialized machine to obtain accurate data.

“Considering MRI method as the example, each session generally lasts 2 h under the supervision and manipulation of professional researchers, and during the scanning process, the subjects are almost fully constrained in the specialized machine in order to obtain accurate data (24).”

• cites [24] Young, 2015, What does an mri scan cost, Healthc Financ ManageNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Additional modules and branches add complexity to registration models.

“Expect the backbone parameters, additional modules and branches also add the complexity of the registration models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

A number of registration methods for medical images based on deep similarity metrics have been studied and have shown great potential in multimodality image registration.

“So far, a number of registration methods for medical image based on deep similarity metrics have been studied and have shown great potential in multimodality image registration.”

May combine two claims.

Image registration and image segmentation are inextricably linked and can help one another.

“image registration and image segmentation are inextricably linked and can help one another.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

From 2013, researchers began to apply deep learning techniques to image registration, and applications of deep models for registration flourished from 2017.

“From 2013, researchers began to apply deep learning techniques to image registration, and the applications of deep models for registration flourished from 2017.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Deep similarity-based models and reinforcement learning-based methods adopt an iterative strategy, which is time-consuming.

“Deep similarity-based models and reinforcement learning-based methods adopt iterative strategy, which is time-consuming.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Public datasets normally contain scans within the brain, lung, and prostate ROIs because common diseases exist in these regions.

“The public datasets normally contain scans within these three ROIs as common diseases exist in these regions.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Image registration neural networks are gradually moving from handling 2D images to 3D or 4D (dynamic) volumes, and converting from supervised methods to unsupervised methods and weakly supervised methods.

“Image registration neural networks are gradually moving from handling 2D images to 3D or 4D (dynamic) volumes, and converting from supervised methods to unsupervised methods and weakly supervised methods.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Diffeomorphic image registration using deep models has caught the attention of researchers due to characteristics like topology preservation and transformation invertibility.

“Due to characteristics like topology preservation and transformation invertibility, diffeomorphic image registration using deep models has caught the attention of researchers.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The learning of large and irregular deformation caused by organ movement (e.g., lung respiratory motion) is still a challenging problem and needs further research.

“The learning of large and irregular deformation caused by organ movement (e.g., lung respiratory motion) is still a challenging problem and needs further research.”

Classified as probably not a substantial claim.

From the point of view of registration efficiency, the training time and network parameters still need promotion, especially for 3D image registration.

“From the point of view of registration efficiency, the training time and network parameters still need promotion, especially for 3D image registration.”

Classified as probably not a substantial claim.

The guarantee for diffeomorphism in learning-based methods remains a challenging problem.

“the guarantee for diffeomorphism in learning-based methods remains a challenging problem.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Deep learning-based methods have shown high computational efficiency and comparable accuracy compared with traditional methods.

“Deep learning-based methods have shown high computational efficiency and comparable accuracy compared with traditional methods.”

Probably not a substantial claim: may not pass S3 (load-bearing).

One strategy to address large deformation is to adopt a multi-stage coarse-to-fine architecture.

“one strategy to address the large deformation is to adopt multi-stage coarse-to-fine architecture.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

More papers implement GAN in diverse registration tasks.

“currently, more papers implement GAN in diverse registration tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The most commonly used technique for tackling this issue is to add the L2 norm on the gradient of the deformation field to regularize the predicted deformation.

“The most commonly used technique for tackling this issue is to add the L2 norm on the gradient of the deformation field to regularize the predicted deformation.”