{"id":"70283299-d30d-4102-bca7-e64324976732","arxiv_id":"2508.21154","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A joint 3D Gaussian-splatting-based reconstruction and registration network registers spine CT to two X-rays with 1.14 mm mean error in 0.82 seconds on a small in-house set.","lead":"This paper presents a deep learning system that reconstructs a 3D bone model from two X-rays and registers it to a CT scan, reporting sub-2 mm accuracy in under a second for spine surgery navigation. It is a potentially useful tool for image-guided surgery, but the evaluation is on a small private dataset with a patient-specific fine-tuning advantage.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Patient-specific third-stage pre-training uses the test fold's CT to synthesize DRRs; the reported mTRE/SR may reflect memorized anatomy rather than generalizable registration.","rationale":"The reader's verdict is CONDITIONAL and mentions 'patient-specific pre-training' as a caveat, but the stated weakest assumption is about two-view reconstruction fidelity (Sec. 2, Table 1). My read agrees that the evaluation is the fragile point, yet the specific load-bearing issue is more concrete: the third-stage pre-training appears to use the held-out fold's CT volume to generate training DRRs, making the test patient's anatomy part of the training data. This is a direct protocol leak that can inflate the registration numbers independently of reconstruction quality. It is not an attack on the method's internal consistency; the architecture is plausible and the ablations are informative. But the central claim of 'state-of-the-art' accuracy is only meaningful if the comparison is fair. Since the paper's own ablation shows that the third stage contributes a large improvement (1.36→1.14 mm, 80→93.33% SR), and that stage is exactly the one with target-CT contamination, the result should be treated as conditional on verification with a clean protocol. I therefore recommend keeping the verdict CONDITIONAL, with the added condition that the target-CT DRR stage be either removed or explicitly justified as part of an intended patient-specific adaptation workflow, with comparisons to baselines that are given the same adaptation opportunity.","tokens_in":7674,"tokens_out":4950,"duration_ms":54951,"concrete_test":"Re-run the full evaluation with the third-stage patient-specific pre-training disabled, i.e., remove all DRR/CT pairs synthesized from the held-out fold's CT volumes from every training stage, and report per-fold mTRE, CR, SR on real X-rays. If the PFULL result (mTRE 1.14 mm, SR 93.33%) degrades substantially toward the P2 result (1.36 mm, 80.00%), the headline accuracy is not evidence of generalizable registration. As a second check, train stage 3 using only DRRs synthesized from the training folds' CTs (a clean version of the current protocol) and compare against the reported numbers.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (mTRE 1.14 mm, SR 93.33%, 0.82 s) rests on the evaluation protocol in Sec. 3.2. The three-stage pre-training includes a final stage 'on the target CT volume with DRRs': 'For each fold, the validation set of preoperative CT volumes generated 1,800 biplanar DRR pairs and 600 CT pairs for the third stage.' Under the five-fold cross-validation described immediately before, the 'validation set' is the held-out fold. This means, for each test patient, the exact preoperative CT volume used as the registration target was also used to generate synthetic training pairs before inference. The registration module is therefore trained on the same anatomy it must register during testing, giving it access to a patient-specific shape and density prior. Table 2 isolates the effect: moving from P2 (two-stage pre-training) to PFULL (including the target-CT third stage) improves SR from 80.00% to 93.33% and mTRE from 1.36 mm to 1.14 mm. This improvement is the stage most suspect for leakage. Consequently, the reported 'state-of-the-art' comparison against DiffPose, DDGS-CT, and TS-SAR does not establish generalization to an unseen patient's CT; it may only show that the network can memorize the target volume. The method could still be clinically viable if patient-specific pre-training is an intended part of the workflow, but the paper presents five-fold cross-validation as evidence of generalizable performance without disclosing this training-target overlap as a limitation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents RadGS-Reg, a learning-based framework for vertebral-level CT/X-ray registration. The method first reconstructs a 3D Radiative Gaussians (RadGS) volume from biplanar AP/LA X-rays using an image encoder, a counterfactual attention module, an X-ray rasterizer, and a density voxelizer. This reconstructed volume is then registered to a preoperative CT volume by a 3D/3D registration network that regresses a 6-DOF rigid pose. Training is performed in stages: (1) simulated DRRs from the public VERSE dataset, (2) real intraoperative X-rays from an in-house 10-case dataset, and (3) synthetic DRRs and CT pairs generated from the target CT volume of each cross-validation fold. On the in-house dataset the authors report an SSIM of 94.51%, PSNR of 28.80 dB, mTRE of 1.14 mm, success rate of 93.33%, and runtime of 0.82 s, and they claim state-of-the-art performance over existing reconstruction and registration methods.","tokens_in":8103,"tokens_out":3809,"duration_ms":41834,"significance":"If the reported results hold under a clean evaluation protocol, the paper makes a useful contribution: it recasts 2D/3D registration as 3D/3D registration via a learned radiologic Gaussian representation, introduces a counterfactual attention mechanism to handle noisy biplanar X-rays, and demonstrates that joint/sequential training of reconstruction and registration can improve both. The availability of code and the internal consistency of the ablation study are strengths. However, the central significance claim is currently undermined by the evaluation protocol, because the third pre-training stage uses the held-out fold's target CT volume to generate additional training data before inference. This makes the comparison with existing methods unfair and means the reported accuracy may reflect target-specific adaptation rather than generalizable registration accuracy. The small in-house dataset (10 cases, 30 vertebral-level samples) further limits the strength of the conclusions.","major_comments":[{"comment":"The three-stage pre-training includes a stage trained on the target CT volume of the held-out fold: 'For each fold, the validation set of preoperative CT volumes generated 1,800 biplanar DRR pairs and 600 CT pairs for the third stage.' Since the validation set is the test fold, the registration and reconstruction modules are fine-tuned on the exact anatomy they are later asked to register. This is transductive rather than circular (the test X-rays are not used), but it gives RadGS-Reg a patient-specific shape and density prior that the compared baselines do not receive. Table 2 shows that this stage is responsible for a large gain: SR rises from 80.00% (P2) to 93.33% (PFULL) and mTRE falls from 1.36 mm to 1.14 mm. The authors must either re-run the evaluation without the target-CT third stage, or clearly reframe the method as a patient-specific adaptation pipeline and give the baselines","section":"Sec. 3.2 / Table 2"},{"comment":"The in-house evaluation is very small: 10 intraoperative cases and 30 vertebral-level samples, with five-fold cross-validation. Multiple vertebrae from the same patient may be correlated, yet the statistical significance test ('*', p<0.05) is reported without stating the test procedure, whether the samples are treated as independent, or whether correction for multiple comparisons was applied. The paper should report per-patient or per-vertebra error distributions, confidence intervals, and ideally an independent external dataset. Without this, the 1.14 mm mTRE and 93.33% SR are not sufficient evidence that the method generalizes beyond this particular acquisition setup.","section":"Sec. 3.1 / Table 1"},{"comment":"The comparison with DiffPose, DDGS-CT, and TS-SAR does not specify the initialization or pose perturbation range used for these iterative baselines. The proposed method directly regresses the pose, whereas render-and-compare methods are initialization-dependent. If the baselines were initialized far from the true pose or were not given the benefit of a multi-start search, their low success rates may be an artifact of the protocol. The authors should detail the common evaluation conditions, including initial pose distributions, number of restarts, and whether the same target-CT adaptation was applied to any baseline.","section":"Sec. 3.3 / Table 1"}],"minor_comments":[{"comment":"The notation is confusing: the text says 'replacing the factual attention maps¯A with random weights,' but the equation uses A and a counterfactual distribution γ. Please define A, A-bar, and the intervention do(·) precisely.","section":"Eq. (1)"},{"comment":"The text says the reconstruction comparison uses 'Lrec from Eq. (2)', but Eq. (2) is named LCAL-rec and Lrec is only mentioned in the following sentence. Please define Lrec explicitly and state which loss is used for each compared reconstruction method.","section":"Sec. 2, Eq. (2)"},{"comment":"Typo: 'real-scent intraoperative cases' should be 'real-scene' or 'real clinical intraoperative cases'.","section":"Sec. 3.1"},{"comment":"The hardware description reads 'AMD Core EYPC 7R32'; the correct vendor spelling is 'AMD EPYC'. Also, the table caption mentions 'results marked by the dotted line,' but no dotted line appears in the table; please align the text with the table.","section":"Sec. 3.2 / Table 1"},{"comment":"The definition of CR is unclear: 'the acceptable CR interval was set to 5 mm' and Table 1 reports CR as ranges such as '20-25' and '10-15'. Please define capture range precisely and explain how the interval is computed from the reported tests.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern raised by the skeptic is valid and is also apparent from the paper's own description in Sec. 3.2: the third-stage pre-training uses each held-out fold's target CT to generate training pairs, so the headline comparison does not establish generalization to an unseen CT. This is fixable within the manuscript's scope, but it requires either removing that stage from the main evaluation or explicitly presenting the method as a patient-specific adaptation workflow and adapting the baselines accordingly. The small dataset and unclear statistical testing are additional reasons not to accept the paper in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nPunchline: RadGS-Reg is a plausible engineering synthesis—learn a radiative-Gaussians volume from biplanar X-rays, then regress a 6-DOF pose to the CT volume—and the reported numbers (1.14 mm mTRE, 93.33% SR, 0.82 s) would be notable for spine navigation if the evaluation were clean. The architecture itself is not deeply novel: the pieces come from R2-GS, CAL, and a standard pose head. The joint training and three-stage pretraining are the real contributions, and the ablation is internally consistent. Credit is earned for reporting speed, running ablated variants, and linking the code.\n\nThe load-bearing soft spot is exactly what the stress-test note flags. Section 3.2 says the third pretraining stage generates 1,800 DRRs from the validation fold's preoperative CT volumes, and 'validation' there is the held-out fold. So the network gets to train on the test patient's CT, in synthetic projection form, before seeing that patient's real X-rays. Table 2 isolates the cost: moving from two-stage to full three-stage pretraining lifts SR from 80.00% to 93.33% and cuts mTRE from 1.36 to 1.14 mm. A large chunk of the reported superiority lives in that stage, and the baselines are not given the same patient-specific fine-tuning. That makes the SOTA comparison unfair as reported. Is the transductive setup itself a flaw? Not automatically: a real workflow could fine-tune on the preop CT before a case. But the paper presents five-fold CV as evidence of generalization, hides the overlap behind the word 'validation set', and never compares against baselines given the same adaptation. So the headline claim overreaches.\n\nSecondary issues: the in-house set is 10 cases / 30 vertebral-level samples; no external data, no per-fold error bars. The reconstruction comparison (94.51% SSIM) is on the same small, adapted setup. The two-view reconstruction can lean on the encoder's CT prior, which may be legitimate but doesn't strengthen the generalization story. I did not verify the code.\n\nWho is this for: people building CT/X-ray registration for spine navigation will want to read it and borrow the joint-training idea. It deserves a serious referee—the idea is worth testing on a bigger public set with a fair baseline protocol, either removing the third stage for a generalization claim or giving it to all baselines and labeling the method as patient-specific.\n\nMy recommendation: send it to review, but the reviewers should demand the protocol fix and reframing. I would not treat the absolute mTRE/SR values as established yet.\n\nBest,","headline":"Useful idea, honest ablation, but the third-stage pretraining on the test fold's CT is a real evaluation flaw; the headline numbers need a fair baseline comparison before being called SOTA.","tokens_in":8497,"tokens_out":7227,"would_cite":true,"duration_ms":68676,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RadGS-Reg claims that spine CT/X-ray registration is best done in 3D: reconstruct a radiative-Gaussians volume from two X-rays, then regress its 6-DOF pose against the CT—1.14 mm error, 93.33% success, 0.82 s.","keywords":["CT/X-ray registration","2D/3D registration","Radiative Gaussians","3D Gaussian Splatting","biplanar X-ray reconstruction","vertebral-level registration","counterfactual attention","image-guided navigation"],"falsifier":"Run the method on a public multi-center spine dataset with independent ground truth, e.g., pose from an optically tracked C-arm, using vertebrae not seen in training and withholding the patient-specific third pre-training stage; if the success rate under 2 mm falls to the level of the compared methods, then the reported 93.33% depends on per-patient adaptation rather than the joint reconstruction-plus-registration mechanism itself.","tokens_in":7667,"feed_emoji":"🦴","tokens_out":7342,"duration_ms":69891,"temperature":0.7,"pith_summary":"This paper tries to show that the hard 2D/3D problem of matching intraoperative X-rays to a preoperative spine CT becomes easier if you lift it into 3D first: reconstruct a volumetric Radiative Gaussians (RadGS) model from just two X-rays, then register that volume to the CT volume with a learned 6-DOF pose network. It argues that this joint reconstruction-plus-registration framework beats both 'render and compare' iterative methods and earlier reconstruction-based approaches, with reported mean target registration error of 1.14 mm, success rate 93.33% under 2 mm error, and runtime 0.82 s on an in-house dataset. The authors also claim the approach is model-agnostic across backbone networks and benefits from a three-stage pre-training that moves from simulated DRRs to real X-rays to patient-specific data. If true, this would give image-guided spine navigation an accurate, near-real-time way to align a pre-op CT to two live X-rays without iterative DRR rendering.","feed_headline":"Two X-rays align a spine CT to 1.14 mm in 0.82 s","feed_subtitle":"Rebuilds the vertebra in 3D, then registers that volume to the CT—93% land within 2 mm.","key_machinery":"The carrying mechanism is 'Radiative Gaussians' (RadGS), a 3D Gaussian-splatting representation in which each Gaussian carries a central density, position, and covariance; an X-ray rasterizer projects these Gaussians to form a DRR-like image, and a density voxelizer turns them into a volume for registration. The counterfactual attention module quantifies which image regions actually drive the reconstruction by comparing factual predictions with predictions made under randomly intervened attention maps, forcing the encoder to focus on vertebra bodies rather than adjacent structures. The registration module is a straightforward image-encoder plus pose-head that maps the six-DOF transform from","core_discovery":"The central claim is that biplanar X-rays contain enough information to reconstruct a Radiative Gaussians (RadGS) volume of a vertebra, and that this volume—not the original X-rays—should be the object registered to the CT. The reconstruction module encodes the two X-rays, uses counterfactual attention (CAL) to suppress interference from overlapping vertebrae, and outputs Gaussians (density, position, covariance) which are rendered back to X-rays and voxelized into a volume. The registration module concatenates this voxelized RadGS volume with the segmented CT volume and regresses the rigid 6-DOF transform. The two modules are first trained separately, then jointly so that registration error","pith_inferences":["A natural next test is whole-spine or multi-level registration: the current evaluation is vertebral-level, so the method's sensitivity to inter-vertebral pose differences remains open.","Because the third pre-training stage uses the target CT with DRRs, the method implicitly assumes a per-patient calibration step before surgery; a reader might infer that removing this stage would lower accuracy, as the ablation results suggest.","If the reconstructed RadGS volume faithfully encodes geometry, the same volume could be reused for tasks other than registration—e.g., pre-operative planning or automatic verification of C-arm positioning—without extra DRR generation.","A direct extension would be replacing the rigid pose head with a deformation field to handle soft tissue and non-rigid regions, which the paper lists as future work."],"forward_implications":["Vertebral CT/X-ray registration no longer needs to iterate render-and-compare during inference; the 0.82 s runtime means the alignment can be completed within a typical intraoperative pause.","Because the reconstruction is optimization-free and uses only two X-rays, the method could work with standard AP/LA C-arm setups rather than dense-view sweeps.","Joint training ties reconstruction quality to registration accuracy, so improvements in either module should translate into lower target registration error.","The model-agnostic backbone results suggest the pipeline can be re-hosted on different network architectures without redesigning the training scheme.","The three-stage pre-training path suggests simulated data alone is insufficient; real and patient-specific data are needed for clinical-grade accuracy."],"supporting_citations":[{"why":"Supplies the Radiative Gaussians representation, X-ray rasterizer, and density voxelizer that the reconstruction module builds on.","marker":"[26]"},{"why":"Provides counterfactual attention learning, used to focus reconstruction on vertebral regions.","marker":"[22]"},{"why":"Segments vertebrae in CT, enabling vertebral-level registration inputs.","marker":"[19]"},{"why":"Detects vertebral levels in X-rays, providing the X-ray counterpart to CT segmentation.","marker":"[14]"},{"why":"Supplies the public lumbar CT dataset used for first-stage simulated pre-training.","marker":"[1]"},{"why":"Baseline learning-based 2D/3D registration method and source of the geodesic-loss weighting.","marker":"[9]"},{"why":"Baseline 3DGS-based registration method compared in Table 1.","marker":"[8]"},{"why":"Baseline Lift3D-based spine registration method compared in Table 1.","marker":"[28]"},{"why":"Baseline NeRF-based sparse-view reconstruction method compared in the reconstruction experiments.","marker":"[3]"},{"why":"Baseline 3D Gaussian reconstruction method compared in the reconstruction experiments.","marker":"[6]"}],"fun_headline_variants":["Two X-rays rebuild spine in 3D, then align CT to 1.14 mm","Spine CT registration from biplanar X-rays: 1.14 mm, 0.82 s","RadGS-Reg: joint 3D reconstruction and registration for spine CT/X-ray","From two X-rays to CT alignment: RadGS reconstructs then registers","Counterfactual attention aids spine 3D from noisy X-rays for CT match"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"That a vertebra's true 3D density and shape can be recovered well enough from only two X-rays—through the learned reconstruction—that a network finding the pose of that reconstructed volume is solving the same problem as aligning the patient; if two-view reconstruction is ambiguous or biased in depth, the reported sub-2 mm errors do not necessarily transfer to real alignment.","fun_headline_variants_meta":{"raw":{"variants":["Two X-rays rebuild spine in 3D, then align CT to 1.14 mm","Spine CT registration from biplanar X-rays: 1.14 mm, 0.82 s","RadGS-Reg: joint 3D reconstruction and registration for spine CT/X-ray","From two X-rays to CT alignment: RadGS reconstructs then registers","Counterfactual attention aids spine 3D from noisy X-rays for CT match"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000286,"raw_usage":{"total_tokens":1536,"prompt_tokens":777,"completion_tokens":759,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":656}},"tokens_in":521,"tokens_out":759,"duration_ms":7231,"temperature":1.0,"reasoning_tokens":656,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:32:30.212224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a public multi-center spine dataset with independent ground truth, e.g., pose from an optically tracked C-arm, using vertebrae not seen in training and withholding the patient-specific third pre-training stage; if the success rate under 2 mm falls to the level of the compared methods, then the reported 93.33% depends on per-patient adaptation rather than the joint reconstruction-plus-registration mechanism itself.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Provides counterfactual attention learning, used to focus reconstruction on vertebral regions."},{"cited_title":"In: VISIGRAPP (5: VISAPP)","cited_arxiv_id":null,"evidence_quote":"Segments vertebrae in CT, enabling vertebral-level registration inputs."},{"cited_title":"In: 2024 IEEE Inter- national Conference on Bioinformatics and Biomedicine (BIBM)","cited_arxiv_id":null,"evidence_quote":"Detects vertebral levels in X-rays, providing the X-ray counterpart to CT segmentation."},{"cited_title":"Medical Image Analysis73, 102166 (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the public lumbar CT dataset used for first-stage simulated pre-training."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Baseline learning-based 2D/3D registration method and source of the geodesic-loss weighting."},{"cited_title":"Biomedical Signal Processing and Control 95, 106384 (2024)","cited_arxiv_id":null,"evidence_quote":"Baseline Lift3D-based spine registration method compared in Table 1."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Baseline NeRF-based sparse-view reconstruction method compared in the reconstruction experiments."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Baseline 3D Gaussian reconstruction method compared in the reconstruction experiments."}],"review_version":1}