The Impact of Machine Learning on 2D/3D Registration for Image-Guided Interventions: A Systematic Review and Perspective

May combine two claims.

Image-based navigation techniques do not require specialized equipment but rely on traditional intra-operative imaging.

“Image-based navigation techniques do not require specialized equipment but rely on traditional intra-operative imaging that enabled the minimally invasive technique in the first place.”

• cites Nolte et al., 2000 sourcePartly supported

  The Discussion says the system uses widely available C-arm fluoroscopy for intraoperative imaging. But the System components section describes an optoelectronic navigator, tracked instruments, and custom software, so the claim that it requires no specialized equipment is not supported. Page numbers are not provided in our copy. (LLM, conf=0.90)

  Experimental — accuracy not yet measured
• cites Mirota et al., 2011 sourceNot checked — no full text of the source
• cites Leonard et al., 2016 sourceDisputed

  The Introduction and Method describe navigation based on endoscopic images registered to a CT scan, supporting the imaging portion of the claim. But the Method says the endoscope’s motion is tracked with a 6DOF magnetic tracker to estimate reconstruction scale, and the Results describe attaching that tracker to the endoscope. This contradicts the claim that the technique requires no specialized equipment. Page numbers were not available in our copy. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

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

Because image-based navigation techniques are designed to seamlessly integrate into conventional surgical workflows, their use should-in theory-not cause frustration or prolonged procedure times.

“Further, because image-based navigation techniques are designed to seamlessly integrate into conventional surgical workflows, their use should—in theory—not cause frustration or prolonged procedure times (Vercauteren et al., 2019).”

• cites Vercauteren et al., 2019 sourceSupported

May combine two claims.

Hand-crafted similarity metrics are positively correlated with pose differences when perturbations are small, but are generally non-convex and fail to accurately represent pose offsets when perturbations are large.

“While these metrics are positively correlated with pose differences when the perturbations in {K^v,T^xv,ω^D} are small, they are generally non-convex and fail to accurately represent pose offsets when perturbations are large.”

• cites Berger et al., 2016 sourceNot checked — no full text of the source
• cites Maes et al., 1997 sourcePartly supported

  The Introduction says correlation of sparse features such as edges and ridges can have a peaked optimum near registration but be insensitive to misregistration at larger distances. The paper does not establish that hand-crafted metrics are generally non-convex or that they fail to represent large pose offsets; its Discussion reports recovering misregistrations as large as 10 degrees and 40 mm. No page numbers are available in our copy. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

Navigation systems are not widely adopted due to high purchase price despite limited versatility, increased procedural time and cost, and potential for disruptions to surgical workflows due to line-of-sight occlusions or other system complications.

“Unfortunately, navigation systems are not widely adopted due to, among other things, high purchase price despite limited versatility, increased procedural time and cost, and potential for disruptions to surgical workflows due to line-of-sight occlusions or other system complications (Picard et al., 2014; Joskowicz and Hazan, 2016).”

• cites Picard et al., 2014 sourcePartly supported

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

  Experimental — accuracy not yet measured
• cites Joskowicz and Hazan, 2016 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

The systematic review was conducted in accordance with the PRISMA method.

“To this end, we conducted a systematic literature review in accordance with the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) method (Moher et al., 2009) (cf. Figure 2).”

• cites Moher et al., 2009 sourcePartly supported

  The PRISMA statement section identifies the cited work as the source of PRISMA’s checklist and flow diagram. It says their aim is to improve the reporting of systematic reviews, not to provide a method for conducting one. The citation supports the identification of PRISMA but not the description of it as a review-conduct method. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

The error metrics used in the reviewed studies included standard registration pose error, keypoint distances, mean target registration error (mTRE) in 3D, and varied 2D metrics such as reprojection distances, DICE score, and contour differences.

“The error metrics that were used included the standard registration pose error (commonly separated into translational and rotational DoFs), keypoint distances (Chou and Pizer, 2014; Zhang et al., 2020), and the mean target registration error (mTRE) in 3D, and varied other metrics in 2D, such as reprojection distances (Bier et al., 2019), segmentation or overlap DICE score (Zhang et al., 2020), or”

• cites Chou and Pizer, 2014 sourceNot checked — no full text of the source
• cites Zhang et al., 2020 sourceNot checked — no full text of the source
• cites Bier et al., 2019 sourceSupported

May combine two claims.

Only three studies in the review register multiple objects through learning-based approaches, and these independently registered single objects to obtain relative poses.

“We found that only three studies in our review register multiple objects through learning-based approaches (Varnavas et al., 2013, 2015b; Wang et al., 2020), and moreover, these studies independently registered single objects in order to obtain relative poses.”

• cites Varnavas et al., 2013 sourceNot checked — no full text of the source
• cites Varnavas et al., 2015bNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Wang et al., 2020 sourceSupported

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

The authors' group has previously released a dataset of CTs and >350 X-rays across multiple viewpoints of six cadaveric specimens prior to periacetabular osteotomy.

“To this end, our group has previously released a relatively large dataset of CTs and >350 X-rays across multiple viewpoints of six cadaveric specimens prior to undergoing periacetabular osteotomy (cf. Figure 4) (Grupp et al., 2020).”

• cites Grupp et al., 2020 sourcePartly supported

  The Materials and Methods (Simulation Study; page unavailable) reports CT scans from six cadaveric specimens and simulated fluoroscopic images from multiple views. The Cadaver Study describes collected intraoperative fluoroscopy, but our copy does not establish that the authors released a dataset containing more than 350 pre-osteotomy X-rays; it identifies only a supplementary video as available with the paper. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

Although Grupp et al. (2020c) registered multiple objects, the regularization function employed learned landmark annotations from only a single object.

“Although Grupp et al. (2020c) registered multiple objects, the regularization function employed learned landmark annotations from a only a single object.”

• cites Grupp et al., 2020c sourcePartly supported

  The Materials and Methods section describes multi-object registration, supporting the first part of the claim. It describes regularization of pose parameters, not the use of learned landmark annotations; the landmark used to initialize pelvis registration was manually annotated. Page numbers are unavailable in our copy. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Evaluation on synthetic data may include data sampled from a broad range of viewpoints or deformations that are approximately uniformly distributed, whereas real data is not uniformly sampled but clustered around certain viewpoints.

“For example, evaluation on synthetic data may include data sampled from a broad range of viewpoints or deformations that are approximately uniformly distributed. Real data, on the other hand, is not uniformly sampled from such a distribution, but rather, will be clustered around certain viewpoints (see (Grupp et al., 2020c) for a visualization of viewpoints used during a cadaveric surgery vs the”

• cites Grupp et al., 2020c sourcePartly supported

  The Materials and Methods sections describe simulated views generated from a perturbed AP view and random C-arm rotations, and fluoroscopic views collected during cadaver surgeries (pages unavailable in our copy). They support the distinction between simulated and surgical viewpoints, but do not establish that the simulated viewpoints were approximately uniformly distributed or that the real viewpoints clustered around particular angles. The stated uniform sampling in the simulation passage concerns landmark-offset directions, not viewpoints. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites Bier et al., 2019 sourcePartly supported

  The Data generation section describes synthetic X-rays generated across a broad range of viewing angles, and the Synthetic data results evaluate detection across viewing directions. It does not say those angles were sampled approximately uniformly or that real surgical images cluster around particular viewpoints; it also does not address deformations. Page numbers are unavailable in our copy. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

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

Registration uncertainties have the potential to augment existing robotic control methods that rely on intermittent imaging.

“Finally, we note that registration uncertainties have the potential to augment existing robotic control methods which rely on intermittent imaging (Alambeigi et al., 2019).”

• cites Alambeigi et al., 2019 sourcePartly supported

  The Problem Formulation models uncertainty in image-derived position estimates as measurement noise, and the Conclusion proposes extending SCADE to real-time control. The paper also uses intermittent images, but it does not specifically propose using registration uncertainties to augment existing robotic control methods. (LLM, conf=0.79)

  Experimental — accuracy not yet measured

May combine two claims.

Some uncertainty estimation approaches report interval estimates by reformulating the registration objective as a probability distribution and drawing samples, using test-time NN drop-out, or sampling using test-time deformation covariance matrices in a variational autoencoder.

“Some of these approaches report interval estimates, either by reformulating the registration objective function as a probability distribution and drawing samples (Risholm et al., 2013; Le Folgoc et al., 2016; Schultz et al., 2018), approximating the sampling process using test-time NN drop-out (Yang et al., 2017) or sampling using the test-time deformation covariance matrices embedded within a”

• cites Risholm et al., 2013 sourceSupported
• cites Yang et al., 2017 sourcePartly supported

  The Methods section (2.2.2, page unavailable) says dropout remains enabled at test time to sample multiple predictions, and that the resulting deformation-field variance estimates uncertainty. The Results section (3.1, page unavailable) reports a variance-based uncertainty map, not interval estimates. The other sampling approaches in the citing sentence are attributed to other sources. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites Dalca et al., 2019 sourcePartly supported

  Section 2.2 (p. 3) describes a network that outputs deformation-related mean and covariance parameters, and Section 4.4.4 (p. 11) shows sampled velocity fields and their empirical variance. But Section 2.4 (p. 5) says the covariance is not used at test time and uncertainty analysis is beyond the paper’s scope. The source therefore supports covariance-based sampling as a capability, not the attributed reporting of interval estimates from test-time sampling. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

The remaining 48 articles were included for an in-depth full text review, analysis, and data extraction.

“The remaining 48 articles were included for an in-depth full text review, analysis, and data extraction, which was performed by five of the authors (MU, CG, MJ, YH, and RG).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims. The quote only partly states this claim.

The search of PubMed, embase, Business Source Ultimate, and Compendex resulted in 559 records initially, and after removal of duplicates, 495 unique studies were included for screening.

“The initial search resulted in 559 records, and after removal of duplicates, 495 unique studies were included for screening.”

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

1 included paper did not specify the 3D modality.

“Other modality combinations included 2D RGB to 3D CT (2) and 3D MR (1), or did not specify the 3D modality (1).”

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

Many studies are only evaluated on synthetic data, which may have sometimes been generated by sampling the PCA model also used for registration, introducing inverse crime, and may not have paired 3D data for extensive quantitative evaluation.

“This is further emphasized by the fact that many studies are only evaluated on synthetic data (which we fear may have sometimes been generated by sampling the PCA model also used for registration, introducing inverse crime) and may not have paired 3D data for extensive quantitative evaluation.”

May combine two claims.

Most methods in direct parameter regression limit themselves to shape estimation and assume that a global rigid alignment is either performed prior or unnecessary.

“As noted in Section 2.5, most methods here limit themselves to shape estimation and assume that a global rigid alignment is either performed prior or unnecessary.”

May combine two claims.

Two other studies that deal with multiple objects are limited to object detection and inpainting, respectively, and do not report registration results.

“Two other studies that deal with multiple objects, however, are limited to object detection (Doerr et al., 2020) and inpainting (Esfandiari et al., 2021), respectively, and do not report registration results.”

May combine two claims.

The remaining study (Grupp et al., 2020c) performs a 2D segmentation of multiple bones but does not apply additional learning to perform registration.

“The remaining study (Grupp et al., 2020c) performs a 2D segmentation of multiple bones, but does not apply any additional learning to perform the registration.”

May combine two claims.

Mitrovi et al. (2014) compare a registration estimate to known local minima and thresholds to determine success/failure, which worked well but may be limited in practice as the approach seems to assume knowledge of the correct solution.

“Finally, Mitrovi et al. (2014) compare a registration estimate to known local minima and thresholds to determine success/failure, which worked well but may be limited in practice as the approach seems to assume knowledge of the correct solution.”

May combine two claims.

Several methods describe innovative learning-based similarity metrics that better reflect distances in geometric parameter space, while other studies present (semi-)global initialization or regularization techniques which may rely on contextual data.

“Several methods describe innovative learning-based similarity metrics that better reflect distances in geometric parameter space, while other studies present (semi-)global initialization or regularization techniques which may rely on contextual data.”

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

Luo et al. (2019) identify and reject poor quality frames in bronchoscopy.

“Luo et al. (2019) identify and reject poor quality frames in bronchoscopy”

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

Methods directed at cephalometry are identical in methodology to the radiation therapy methods.

“Methods directed at cephalometry (Li et al., 2020; Zhang et al., 2018; Pei et al., 2017) are identical in methodology to the radiation therapy methods.”

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

The means to accomplishing increased capture range are diverse.

“The means to accomplishing this, however, are diverse.”

May combine two claims.

Both Tang and Scalzo (2016) and Neumann et al. (2020) rely on concepts that are limited to sparse objects, such as vessels.

“Both methods rely on concepts that are limited to sparse objects, such as vessels.”

May combine two claims.

Most emphasis was given on increasing the capture range of the registration pipeline and very little, if any, attention is paid to increasing the resolution and precision of these metrics.

“Further, we found that most emphasis was given on increasing the capture range of the registration pipeline and very little, if any, attention is paid to increasing the resolution and precision of these metrics.”

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

Esfandiari et al. (2021) contribute a non-blind image inpainting method using a FCN that seeks to restore background anatomy after image regions corresponding to instruments were identified.

“They contribute a non-blind image inpainting method using a FCN that seeks to restore the background anatomy after image regions corresponding to instruments were identified.”

May combine two claims.

Miao et al. (2016a,b, 2018) and Zheng et al. (2018) use the image residual between target and source in a region of interest around the projected tool model; because only a small part of the target image is used, the initialization must be sufficiently close.

“(Miao et al., 2016a,b, 2018; Zheng et al., 2018). use the image residual between y v and y^v in a region of interest around the projected tool model. Because only a small part of the target image is used, the initialization must be sufficiently close.”

May combine two claims.

Among studies proposing novel image similarity functions, two methods rely on regularization: Grupp et al. detect anatomical landmarks to expand an analytic similarity function with landmark-reprojection constraints, while Francois et al. segment occluding contours to constrain similarity evaluation to salient regions.

“Among those, two methods rely on regularization: Grupp et al. (2020c) detect anatomical landmarks to expand an analytic similarity function with landmark-reprojection constraints to enhance the capture range of an intensity-based strategy, while Francois et al. (2020) segment occluding contours to constrain similarity evaluation to salient regions.”

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

The remaining 14 methods consider rigid parameter regression.

“The remaining 14 methods consider rigid parameter regression”

May combine two claims.

Cephalometry methods rely on population-level models, different from radiation therapy and angiography methods which rely on patient-specific models built pre- and intra-operatively, respectively.

“Consequently, these methods rely on population-level models and are thus different from methods used for radiation therapy (Chou et al., 2013; Chou and Pizer, 2013, 2014; Zhao et al., 2014; Foote et al., 2019) and angiography (Brost et al., 2012; Baka et al., 2015), which rely on patient-specific models that are built pre- and intra-operatively, respectively.”

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

Neumann et al. (2020) regress the disparity between corresponding points in source and target images to quantify image dissimilarity.

“Neumann et al. (2020) regress the disparity between corresponding points in source and target images to quantify image dissimilarity”

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

Liao et al. (2019) train a network to establish keypoint correspondences between source and target images.

“To this end, Liao et al. (2019) train a network to establish keypoint correspondences between the source and target images.”

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

While most methods rely on PCA to condense deformable motion parametrizations, it has also been found useful to identify and focus on the primary modes of variation in rigid registration.

“It is worth mentioning that, while most methods rely on PCA to condense deformable motion parametrizations, it has also been found useful to identify and focus on the primary modes of variation in rigid registration (Brost et al., 2012).”

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

Toth et al. (2018) extract features using a CNN from both source and target image independently and then concatenate them for pose regression using fully connected layers.

“Toth et al. (2018) extract features using a CNN from both source and target image independently and then concatenate them for pose regression using fully connected layers.”

May combine two claims.

Gu et al. (2020) and Gao et al. (2020b) estimate the geodesic in Riemannian tangent space between source and target camera poses, which in an ideal case results in a convex similarity function.

“To this end, Gu et al. (2020) and Gao et al. (2020b) estimate the geodesic in Riemannian tangent space between the source and target camera poses, which in an ideal case results in a convex similarity function.”

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

Methods that rely on the target image only are relatively straight-forward and generally train a standard feed-forward CNN architecture to regress pose on large datasets comprising multiple independent objects or anatomies.

“Methods that rely on the target image only are relatively straight-forward and generally train a standard feed-forward CNN architecture to regress pose on large datasets comprising of multiple independent objects or anatomies.”

May combine two claims.

Some verification studies, such as Varnavas et al. (2015b) and Wu et al. (2016), are mature and were demonstrated to work well on comparably large amounts of real clinical data; however, these methods are not general purpose as they rely on registration of multiple objects and availability of two orthogonal views, respectively.

“Some studies included in this theme stand out, in that they are certainly mature and were demonstrated to work well on comparably large amounts of real clinical data, such as (Varnavas et al., 2015b) and (Wu et al., 2016); unfortunately however, these methods are not general purpose as they rely on the registration of multiple objects and the availability of two orthogonal views, respectively.”

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

The fully differentiable pipeline enables both end-to-end learning of Gyθy(y), Gxθx(x) and/or SθS(⋅,⋅) during training as well as analytic gradient-based optimization using backpropagation during application.

“This enables both, end-to-end learning of Gyθy(y),Gxθx(x) and/or SθS(⋅,⋅) during training as well as analytic gradient-based optimization using backpropagation during application.”

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

All methods for landmark detection rely on deep CNNs.

“While all methods for landmark detection rely on deep CNNs”

May combine two claims.

All methods included in this review are limited to registration of a single object, and it remains unclear how these methods would apply to multiple objects, in part because of the combinatorial explosion of relative poses.

“We found that all methods included in this review are limited to registration of a single object and it remains unclear how these methods would apply to multiple objects, in part, because of the combinatorial explosion of relative poses.”

May combine two claims.

PCA is by far the most prevalent method for representation learning and is used in all but one study.

“PCA is by far the most prevalent method for representation learning and is used in all but one study.”

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

In lower dimensional spaces, e.g., the six parameters of a rigid transformation or the principal components of a PCA model, direct pose regression from the target and/or source images is a clearly established line of research.

“In those lower dimensional spaces, e.g., the six parameters of a rigid transformations or the principal components of a PCA model, direct pose regression from the target and/or source images is a clearly established line of research.”

May combine two claims.

None of the studies reviewed here describes a dedicated effort to establish a canonical reference frame, suggesting that those methods will eventually hit a performance ceiling determined by misalignment within the reference coordinate systems in the training data.

“Unfortunately, none of the studies reviewed here describes a dedicated effort to establish such a canonical reference frame, suggesting that those methods will eventually hit a performance ceiling that is determined by the misalignment within the reference coordinate systems in the training data.”

May combine two claims.

Representation learning techniques, currently dominated by PCA, are a clearly established tool to reduce the dimensionality of deformable 2D/3D registration problems, and may even be useful for rigid alignment.

“We found that representation learning techniques, currently dominated by PCA, are a clearly established tool to reduce the dimensionality of deformable 2D/3D registration problems, and may even be useful for rigid alignment.”

May combine two claims.

While several methods report an overall improved registration performance when using their novel, machine learning-enhanced algorithms, no method was registered with a particular focus on reducing ambiguity.

“While several methods report an overall improved registration performance when using their novel, machine learning-enhanced algorithms, we did not register any method with a particular focus on reducing ambiguity.”

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

Doerr et al. (2020) describe a deep learning-based alternative where the Fast R-CNN object detector is re-trained to return bounding boxes of 30 different screw types in varied poses.

“Doerr et al. (2020) describe a deep learning-based alternative, where the Fast R-CNN object detector is re-trained to return bounding boxes of 30 different screw types in varied poses.”

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

The review noted certain trends and recurring shortcomings in the literature.

“Despite this positive prospect on the utility of machine learning for 2D/3D registration, we noted certain trends and recurring shortcomings in our review that we will discuss next.”

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

The more obvious examples of tasks not currently performed in clinical practice are visual servoing of novel robotic surgery platforms or robotic imaging paradigms that alter how data is acquired.

“The more obvious examples are visual servoing of novel robotic surgery platforms (Gao et al., 2019) or robotic imaging paradigms that alter how data is acquired (Zaech et al., 2019; Thies et al., 2020).”

May combine two claims.

Of 32 studies describing deep learning-based methods, 24 trained on synthetic data, 7 on real data, and 1 used a pre-trained network without training.

“Indeed, of the 32 studies that describe deep learning-based methods, 24 trained on synthetic data (seven trained on real data, and one did not train at all but used a pre-trained network).”

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

There are no standard routines to define successful registrations.

“There are no standard routines to define the successful registrations.”

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

There is a lack of universally accepted reporting guidelines, error metrics, and datasets in 2D/3D registration.

“However, we also believe that the lack of universally accepted reporting guidelines, error metrics, and datasets is a severe shortcoming that has unfortunate consequences, such as a high risk of duplicated efforts and non-interpretable performance reporting.”

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

Most code-bases and datasets are kept private, which inhibits reproduction because re-implementation is prone to biased conclusions.

“Further, most code-bases and more importantly datasets are kept private which inhibits reproduction, since re-implementation is particularly prone to biased conclusions.”

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

There is a lack of public benchmarks for 2D/3D registration and related problems.

“While a too strong focus on “winning benchmarks” is certainly detrimental to the creativity of novel approaches, we feel that the lack of public benchmarks for 2D/3D registration and related problems is perhaps an even greater worry.”

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

An issue that has become more prevalent with machine learning methods trained on synthetic data is the lack of substantial results on clinically relevant, real data.

“An issue that appears to have become more prevalent with the introduction of machine learning methods that are developed and trained on synthetic data is the lack of substantial results on clinically relevant, real data.”

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

Most of the studies of machine learning-based improvements to image-based 2D/3D registration are performed "on the benchtop."

“most of those studies are performed “on the benchtop.””

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

Contextualization of data, representation learning to reduce problem dimensionality, similarity modeling for increased capture range, direct pose regression to avoid iterative optimization, and confidence assessment are all well established research thrusts geared towards developing automated registration pipelines.

“Contextualization of data, representation learning to reduce problem dimensionality, similarity modeling for increased capture range, direct pose regression to avoid iterative optimization, as well as confidence assessment are all well established research thrusts, which are geared towards developing automated registration pipelines.”

May combine two claims.

These challenges already exist for rigid 2D/3D registration and are further exacerbated in rigid plus deformable registration settings.

“These challenges already exist for rigid 2D/3D registration and are further exacerbated in rigid plus deformable registration settings.”

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

Surgical navigation promotes reproducibly good patient outcomes, and when combined with robotic assistance systems, may enable novel treatment options and improved techniques.

“Surgical navigation promotes reproducibly good patient outcomes, and when combined with robotic assistance systems, may enable novel treatment options and improved techniques (van der List et al., 2016).”

• cites van der List et al., 2016 sourceNot checked — no full text of the source

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

Image-based 2D/3D registration is a mature technique, but its transition from the bench to the bedside has been restrained by challenges including brittleness with respect to optimization objective, hyperparameter selection, and initialization, difficulties in dealing with inconsistencies or multiple objects, and limited single-view performance.

“While image-based 2D/3D registration is a mature technique, its transition from the bench to the bedside has been restrained by well-known challenges, including brittleness with respect to optimization objective, hyperparameter selection, and initialization, difficulties in dealing with inconsistencies or multiple objects, and limited single-view performance.”

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

Even in the simplest case, 2D/3D registration describes a non-convex optimization problem with at least six degrees of freedom to describe a rigid body transform.

“Even in the simplest case, 2D/3D registration describes a non-convex optimization problem with at least six DoFs to describe a rigid body transform.”

Classified as probably not a substantial claim.

Machine learning methods optimize parametric functions on training data such that the model learns to approximate the desired mapping between input and output variables.

“Rather, they optimize parametric functions, such as convolutional NNs (CNNs), on training data such that the model learns to approximate the desired mapping between input and output variables.”

May combine two claims.

Visual inspection of the 2D overlay of projected 3D data to verify registration is neither quantitative nor scalable because it is based on human intervention.

“Perhaps the most straightforward way of verifying a registration result is to visually inspect the 2D overlay of the projected 3D data—this approach, however, is neither quantitative nor does it scale since it is based on human intervention.”

May combine two claims.

The search did not use modality specific terms, such as video to CT registration, which may have excluded some manuscripts that focus on endoscopy or other RGB camera-based modalities.

“We also did not use modality specific terms, such as “video to CT registration”, which may have excluded some manuscripts that focus on endoscopy or other RGB camera-based modalities.”

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

The list of included articles may not be exhaustive.

“Despite our efforts to broaden the search terms regarding 2D/3D registration, we acknowledge that the list may not be exhaustive.”

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

Estimating good initial parameters is commonly achieved using manual interaction, which is cumbersome and time consuming, or neglected altogether in research papers.

“Estimating good initial parameters is commonly achieved using some manual interaction, which is cumbersome and time consuming, or—in research papers—neglected all-together.”

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

Every one of the 48 articles was analyzed and coded by at least two of the five authors, and one author merged the individual reports into a final consensus document.

“Every one of the 48 articles was analyzed and coded by at least two of the five authors and one author (MU) merged the individual reports into a final consensus document.”

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

Image-based navigation techniques do not introduce economic trade-offs.

“Therefore, these techniques do not introduce economic trade-offs.”

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

Limitations regarding cost particularly inhibit the adoption of surgical navigation systems in geographical areas with less specialized healthcare providers with lower volumes for any procedure.

“While frustrations induced by workflow disruption affect every operator equally, the aforementioned limitations regarding cost particularly inhibit the adoption of surgical navigation systems in geographical areas with less specialized healthcare providers with lower volumes for any procedure; areas where routine use of navigation would perhaps be most impactful.”

May combine two claims.

The analysis is limited to records that appeared after January 2012, which pre-dates the onset of the ongoing surge of interest in learning-based image processing.

“Because we are primarily interested in surveying the impact of contemporary machine learning techniques, such as deep CNNs, on 2D/3D registration, we limit our analysis to records that appeared after January 2012, which pre-dates the onset of the ongoing surge of interest in learning-based image processing (Krizhevsky et al., 2012).”

• cites Krizhevsky et al., 2012Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

Publication bias may have resulted in the exclusion of works relevant to this review.

“Publication bias may have resulted in the exclusion of works relevant to this review.”

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

The terminology used in the search may have resulted in the exclusion of relevant work from the general computer vision literature.

“The terminology used in the search may also have resulted in the exclusion of relevant work from the general computer vision literature.”

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

The capture range magnitude depends on the similarity function and the optimizer, and can be stated quantitatively as the mean target registration error between 3D keypoints at initialization.

“Its magnitude depends, among other things, on the similarity function as well as the optimizer, and it can be stated quantitatively as the mean target registration error (mTRE) between 3D keypoints at initialization1.”

• cites Markelj et al., 2012 sourceNot checked — no full text of the source
• cites Esteban et al., 2019 sourceNot checked — no full text of the source

May combine two claims.

Although statistical modeling techniques exist to limit the parameter search space, registration accuracy and sensitivity to key features remain an area of concern.

“Although statistical modeling techniques exist to limit the parameter search space, the registration accuracy and sensitivity to key features remain an area of concern (Zhu et al., 2021).”

• cites Zhu et al., 2021 sourceNot checked — no full text of the source

May combine two claims.

2D/3D registration is shackled to the bench top because several hard open challenges inhibit its widespread adoption.

“Rather, it is shackled to the bench top because several hard open challenges inhibit its widespread adoption.”

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

Analytical solutions are likely inadequate considering the complexity, variability, and high-dimensionality of generic 2D/3D registration problems.

“One reason these challenges persist today is that analytical solutions are likely inadequate considering the complexity, variability, and high-dimensionality of generic 2D/3D registration problems.”

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

The speed of direct pose regression is particularly appealing for traditionally time critical applications, such as tumor tracking in radiation therapy.

“This is particularly appealing for traditionally time critical applications, such as tumor tracking in radiation therapy.”

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

While fully differentiable 2D/3D registration methods are in a development stage, their flexibility may prove a great strength in developing solutions that meet clinical needs.

“While these methods are in a development stage, their flexibility may prove a great strength in developing solutions that meet clinical needs.”

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

Pose regression directly from images is appealing because it may result in substantially faster convergence, potentially with a single forward pass of a CNN.

“Pose regression directly from images is appealing because it may result in substantially faster convergence, potentially with a single forward pass of a CNN.”

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

2D/3D registration, especially in deformable scenarios, suffers from high dimensional parameter spaces and changes in any one parameter are not easily resolved due to limited information which creates ambiguity.

“As highlighted in Section 1.2, 2D/3D registration, especially in deformable scenarios, suffers from high dimensional parameter spaces and changes in any one of those parameters are not easily resolved due to limited information which creates ambiguity.”

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

For optimization-based 2D/3D registration algorithms, the cost function or similarity metric is among the most important components since it determines parameter updates.

“As we established in Section 1.2 for optimization-based 2D/3D registration algorithms, the cost function—or similarity metric—S(⋅, ⋅) is among the most important components since it will determine parameter updates.”

May combine two claims.

Methods that solely rely on the target image never involve neither the 3D data nor the source images created from it; this may be problematic because it is unclear how these methods would verify that the specific 2D/3D registration data satisfy the canonical coordinate frame assumption.

“Further, methods that solely rely on the target image never involve neither the 3D data nor the source images created from it. This may be problematic, because it is unclear how these methods would verify that this specific 2D/3D registration data satisfy, among other things, the canonical coordinate frame assumption.”

May combine two claims.

The demonstrated improvements are generally significant, suggesting huge potential for machine learning in regards to the narrow capture range challenge.

“The demonstrated improvements are generally significant, suggesting huge potential for machine learning in regards to this particular challenge.”

May combine two claims.

Except for few exceptional cases with highly specialized instrumentation, assumptions around rigid alignment seem unjustified, which may suggest that performance estimates need to be interpreted with care.

“Except for few exceptional cases with highly specialized instrumentation, for example (Chou and Pizer, 2013) that relied on on-board CBCT imaging, assumptions around rigid alignment seem to be unjustified, which may suggest that the performance estimates need to be interpreted with care.”

May combine two claims.

Image-based navigation techniques as well as visual servoing of surgical robots will also require high-precision 2D/3D registration at video frame-rates.

“This is because image-based navigation techniques (Sugano, 2003; Hummel et al., 2008; Tucker et al., 2018) as well as visual servoing of surgical robots (Yi et al., 2018; Gao et al., 2019; Unberath et al., 2019) will also require high-precision 2D/3D registration at video frame-rates.”

• cites Sugano, 2003 sourceNot checked — no full text of the source
• cites Hummel et al., 2008 sourceNot checked — no full text of the source
• cites Tucker et al., 2018 sourceNot checked — no full text of the source

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

Most commonly used metrics fail to accurately represent the distances in geometric parameter space that generate the mismatch between the current observations.

“It is well known that most commonly used metrics fail to accurately represent the distances in geometric parameter space that generate the mismatch between the current observations.”

May combine two claims.

Contextualization, such as landmark detection and segmentation, can mimic user input to enable novel paradigms for initialization while also providing a clear interface for human-computer interaction.

“Contextualization, such as landmark detection and segmentation, can mimic user input to enable novel paradigms for initialization while also providing a clear interface for human-computer interaction”

May combine two claims.

Direct pose regression approaches supersede optimization with a few forward passes of a machine learning model, which makes them comparably fast.

“These approaches supersede optimization with a few forward passes of a machine learning model, which makes them comparably fast.”

May combine two claims.

Methods that regress pose between target and source avoid the canonical frame concern because poses are relative rather than absolute, but similarly to conventional techniques, they require an initialization.

“Methods that regress pose between y v and y^v avoid the aforementioned concern because poses are relative rather than absolute, however similarly to conventional techniques, they require an initialization.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The introduction of machine learning to contextualize 2D and 3D data enables novel techniques that substantially expand the tools for designing 2D/3D registration algorithms and are likely to become impactful.

“It is undeniable that the introduction of machine learning to contextualize 2D and 3D data enables novel techniques that quite substantially expand the tools one may rely on when designing a 2D/3D registration algorithm, and as such, are likely to become impactful.”

May combine two claims.

The challenge of defining a canonical 3D coordinate system is mitigated for applications that consider an instrument or tool, because such a canonical system can be readily defined; however, establishing this reference frame is considerably more effortful for patient anatomy and may require (group-wise) 3D/3D registrations.

“This challenge is mitigated for applications that consider an instrument or tool, because such a canonical system can be readily defined; however, establishing this reference frame is considerably more effortful for patient anatomy and may require (group-wise) 3D/3D registrations.”

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

Compelling evidence will need to be provided on potential sources and extent of algorithmic bias and reliable estimates of real-world performance for regulatory approval.

“Then, compelling evidence will need to be provided on potential sources and extent of algorithmic bias as well as reliable estimates of real world performance (US Food and Drug Administration, 2021).”

• cites US Food and Drug Administration, 2021Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

Experimental conditions including ground truth targets are perfectly known for simulation but much harder to obtain for clinical or cadaveric data.

“While experimental conditions, including ground truth targets, are perfectly known for simulation, they are much harder to obtain for clinical or cadaveric data.”

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

Even for the simple opportunity-cost case to succeed, many conditions must be met including ethical review board approval, digital medicine infrastructure, and methods to reliably annotate the data.

“It should be noted, however, that even for this “simple” case to succeed, many conditions must be met including ethical review board approval, digital medicine infrastructure, and methods to reliably annotate the data.”

May combine two claims.

These challenges restrict research and inhibit translational efforts because we cannot reliably understand whether a specific 2D/3D registration problem is solved or what the open problems are.

“These challenges clearly restrict research, but even more dramatically inhibit translational efforts, simply because we cannot reliably understand whether a specific 2D/3D registration problem should be considered “solved” or what the open problems are.”

May combine two claims.

Contemporary deep learning techniques can map complex functions from high-dimensional input but require heavily over-parameterized models and large amounts of data.

“Contemporary deep learning-based techniques enable the mapping of very complex functions directly from high-dimensional input data at the cost of heavily over-parameterized models that require as much data as possible to learn sensible associations.”

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

Not all evaluation paradigms available on synthetic data can be readily transferred to clinical data, but the reverse is not true.

“While it is evident that not all evaluation paradigms that are easily available on synthetic data can be readily transferred to clinical data, the reverse is not true.”

May combine two claims.

A multiple object registration strategy that simultaneously solves for one object's pose relative to the imaging device and the relative pose between objects partially compounds motion and eliminates conflicting depth estimates.

“A multiple object registration strategy could alternatively parameterize the problem to simultaneously solve for the pose of one object with respect to the imaging device and also for the relative pose between the objects. This approach partially compounds motion of the objects during registration and completely eliminates the possibility of conflicting depth estimates.”

Classified as probably not a substantial claim.

Image-guidance systems often need to process and report relative poses between several objects of interest, such as bones, organs, and surgical instruments.

“Image-guidance systems often need to process and report information regarding the relative poses between several objects of interest, such as bones, organs and surgical instruments.”

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

This practice is undesirable because synthetic data experiments cannot serve as a baseline and true quantities of interest remain unknown.

“Clearly, this practice is undesirable because 1) synthetic data experiments now cannot serve as a baseline (since they use different metrics and are thus incomparable), and 2) the true quantities of interest remain unknown (for example, a 3D mTRE is more informative than a 2D reprojection TRE since it can adequately resolve depth).”

Probably not a substantial claim: may not pass S2 (checkable).

In 2D/3D deformable registration for tumor tracking during radiation therapy, accurately recovering 3D tumor shape and position is more relevant than Euclidean distance between deformation field parameters.

“For example, in 2D/3D deformable registration for tumor tracking during radiation therapy, accurately recovering the 3D tumor shape and position (quantified well using, e.g., the DICE score of true and estimated 3D position over time) is much more relevant than a Euclidean distance between the deformation field parameters, which would describe irrelevant errors far from the region of interest.”

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

Increasing the rate of sharing code and data will likely advance research because baselines become transparent and clearly defined.

“Increasing the rate with which we share code and data will likely result in research that truly advances the contemporary capabilities since the baselines are transparent and much more clearly defined.”

Probably not a substantial claim: may not pass S2 (checkable).

Data synthesis is an important idea that enables research on creative approaches advancing 2D/3D registration.

“It is evident that data synthesis is an important idea that enables research on creative approaches that contribute to the advancement of 2D/3D registration.”

May combine two claims.

Although representation learning methods may function with multiple objects without modification, better performance is usually obtained by adding structure to account for known spatial relationships between objects.

“Although representation learning methods may function in the presence of multiple objects without additional modification, better performance is usually obtained by adding structure to account for the known spatial relationships between objects (Yokota et al., 2013).”

• cites Yokota et al., 2013Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

Updating an intensity-based registration framework to compute Eq. 3 instead of Eq. 1 is relatively straightforward from an implementation perspective.

“Indeed, updating an intensity-based registration framework to compute (3) instead of (Eq. 1) is relatively straight forward from an implementation perspective: compute N synthetic radiographs instead of one and perform a pixel-wise sum of the synthetic images together before calculating the image similarity metric, S(⋅, ⋅).”

• AssumptionThe authors state this as an assumption: The 2D view modality is X-ray.

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

An omnipresent concern in the development of machine learning-based components for 2D/3D registration is the availability of or access to large amounts of relevant data.

“An omnipresent concern in the development of machine learning-based components for 2D/3D registration, but everywhere really, is the availability of or access to large amounts of relevant data.”

May combine two claims.

Trained machine learning algorithms approximate the target function only on a compact domain, and their behavior outside this domain is unspecified.

“Trained machine learning algorithms approximate the target function only on a compact domain (Zhou, 2020), and their behaviour outside this domain is unspecified.”

• cites Zhou, 2020 sourceNot checked — no full text of the source

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

The wealth of evaluation strategies and metrics can be partly attributed to different clinical applications necessitating different conditions.

“The wealth of evaluation strategies and metrics can, in part, be attributed to the fact that different clinical applications necessitate different conditions to be met.”

May combine two claims.

Training on synthetic data has clear advantages because large datasets and corresponding annotations can be generated with relatively little effort, and rigid pose and deformation parameters are perfectly known by design, creating an unbiased learning target.

“Training on synthetic data has clear advantages because large datasets and corresponding annotations can be generated with relatively little effort. In addition, rigid pose and deformation parameters are perfectly known by design thus creating an unbiased learning target.”

Probably not a substantial claim: may not pass S2 (checkable).

The lack of new multiple object registration learning-based strategies is indicative of substantial challenges in their development, rather than a perceived lack of importance.

“We believe that the lack of new multiple object registration learning-based strategies is indicative of the substantial challenges involved with their development, rather than any perceived lack of the problem’s importance by the community.”

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

The recent interest in deep learning has brought a move towards open science, with sharing of source code and public benchmarks allowing fair comparisons.

“The recent interest in deep learning has brought about a transformational move towards open science, where large parts of the community share their source code publicly and demonstrate the performance of their solutions on public benchmarks, which allows for fair comparisons.”

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

In some cases, the data problem amounts to a simple opportunity cost, e.g., for automation of manually performed tasks such as landmark detection.

“In some cases, the data problem amounts to a simple opportunity cost, e.g., for automation of manually performed tasks such as landmark detection.”

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

Representation learning methods have the potential to reconstruct bony anatomy of joints from a sparse number of 2D views, which could enable navigation to a 3D anatomical model with reduced radiation exposure.

“As representation learning methods have the potential to reconstruct the bony anatomy of joints from a sparse number of 2D views, these could potentially become very popular by enabling navigation to a 3D anatomical model with substantially reduced radiation exposure.”

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

Foreign objects such as screws, K-wires, retractors, and osteotomes frequently confound traditional registration methods, but their location or poses often have clinical relevance.

“Although foreign objects such as screws, K-wires, retractors, osteotomes, etc. frequently confound traditional registration methods, their location or poses often have clinical relevance.”

May combine two claims.

Retaining the angled osteotome as a static object during the posterior cut helps guide the flat osteotome but does not guarantee that the acetabular fragment has been freed from the pelvis.

“Figure 4 (c) demonstrates that retaining the angled osteotome as a static object in the field of view during performance of the posterior cut helps to guide the flat osteotome along the desired trajectory, but does not guarantee that the acetabular fragment has in fact been freed from the pelvis.”

May combine two claims.

Specific machine learning models are not only suitable for one specific task or anatomy; pose or PCA component regression techniques are largely identical across clinical targets.

“This is not because specific machine learning models are only suitable for one specific task or anatomy; on the contrary, e.g., pose or PCA component regression techniques are largely identical across clinical targets.”

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

A navigation system capable of accurately tracking the osteotome with respect to anatomy using a series of oblique views would eliminate the need for additional confirmation views, potentially reduce operative time, decrease radiation exposure, and reduce breaches into sensitive anatomy.

“A navigation system capable of accurately tracking the osteotome with respect to the anatomy using a series of oblique views would eliminate the need for additional, “confirmation,” views, potentially reduce operative time, decrease radiation exposure to the patient and clinical team and reduce the frequency of breaches into sensitive anatomy.”

May combine two claims.

Methods like Varnavas et al. (2013; 2015b) rely on patient- and object-specific training sets, and therefore avoid the task-specificity that comes with one-time training.

“Methods like (Varnavas et al., 2013; 2015b) rely on patient- and object-specific training sets, and therefore, avoid the task-specificity that comes with one-time training.”

• cites Varnavas et al., 2013; 2015bNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

In osteotome navigation, oblique views help distinguish the osteotome tip from the acetabulum, but lateral orientation is challenging to interpret manually and may need confirmation with subsequent views at substantially different orientations.

“Although the oblique views shown in Figure 4 1) and (c) help ensure that the osteotome tip is distinguishable from the acetabulum, the lateral orientation of the tool is challenging to interpret manually and may need to be confirmed by collecting subsequent fluoroscopic views at substantially different orientations, such as the approximate anterior-posterior (AP) view shown in Figure 4 (b).”

May combine two claims.

The introduction of machine learning to the 2D/3D registration pipeline has taken away the generality of traditional algorithmic registration, and machine learning methods have become substantially more specialized.

“The introduction of machine learning to the 2D/3D pipeline, however, has in a way taken away this generality and methods have become substantially more specialized.”

• AssumptionRests on an assumption: Machine learning models have a compact domain assumption that confines solutions to the data domain they were developed on.

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

There is likely still room for improvement of the local/global minima classification problem.

“Although we have mostly advocated for the development of new uncertainty quantification methods, there is likely still room for improvement of the local/global minima classification problem.”

Classified as probably not a substantial claim.

Registration strategies attempt to find singular solutions, or point estimates, that best minimize the objective function.

“This is perhaps easiest to recognize by first revisiting (1) and noting that registration strategies attempt to find singular solutions, or point estimates, which best minimize the appropriate objective function.”

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

An additional challenge in osteotomy cases is determining when all osteotomies are complete and the new bone fragment is completely free from its parent.

“An additional challenge with osteotomy cases is the ability to determine when all osteotomies are complete and the new bone fragment is completely free from its parent.”

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

Deep learning has accelerated and diversified the contributions to the field of image-based 2D/3D registration.

“and deep learning has only accelerated and diversified the contributions to the field.”

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

Due to the interventional nature of most 2D/3D registration applications, care needs to be taken to ensure program runtimes are compatible with intra-operative workflows.

“Due to the interventional nature of most 2D/3D registration applications, care needs to be taken in order to ensure that program runtimes are compatible with intra-operative workflows.”

May combine two claims.

Despite matching random points, Liao et al.'s method is likely still scene-specific because matching is performed using a very deep CNN with very large receptive field which may have learned to exploit the global context of the random keypoints, which would not be preserved in different anatomy.

“However, despite matching random points, this method is likely still scene-specific because matching is performed using a very deep CNN with very large receptive field which may have learned to exploit the global context of the random keypoints, which would not be preserved in different anatomy.”

• cites Liao et al., 2019 sourceNot checked — no full text of the source

May combine two claims.

Frequent properties of unsolved multiple object registration problems include misleading views with substantial 2D overlap, varying numbers of surgical instruments, and objects that dynamically change shape or split into new objects.

“Some frequent properties of these unsolved problems are misleading views with several objects having substantial overlap in 2D, the potential for a varying number of surgical instruments to be present in a view and the existence of objects which are dynamically changing shape or possibly “splitting” into several new objects.”

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

Given contemporary GPUs and learning frameworks, it should be feasible to extend Wu et al. (2016) by passing densely sampled regions of the objective function through a NN to distinguish local and global minima.

“Given the computational capacity of contemporary GPUs and sophistication of learning frameworks, it should be feasible to extend the approach of Wu et al. (2016) and pass densely sampled regions of the objective function through a NN, relying on the learning process to extract optimal features for distinguishing local and global minimal.”

• cites Wu et al., 2016Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

Learning-based methods show great promise and have supplanted traditional methods in many aspects.

“Even though learning-based methods show great promise and have supplanted traditional methods in many aspects”