{"id":"9ba42817-2826-4a52-b9c8-351dedf09a02","arxiv_id":"2508.14708","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A rule-based pipeline estimates vertebral orientation and extracts landmark points from MRI for biomechanical spine models.","lead":"This paper adapts a CT-based rule-based method to extract anatomical landmarks from spine MRI scans, adding a new step that estimates each vertebra's orientation before placing the landmarks. The goal is to enable radiation-free, subject-specific biomechanical simulations of the spine, and the code is released open source.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated hand-picked raycasting directions (Sec. 2.2) are the load-bearing weak point: process POIs for the MBS depend on fixed coefficients not derived or tested across levels/pathologies, so the 'subpixel-accurate' MRI claim is not established.","rationale":"The reader's weakest_assumption identifies exactly this: the hand-picked raycasting coefficients in Section 2.2 are the most load-bearing unvalidated component of the method. The paper's novelty is the adaptation to MRI, and the process endpoints are critical for the stated application to multibody spine models. If those directions are wrong for a given vertebral level, pathology, or scan protocol, every downstream POI for the transverse and spinous processes is systematically shifted, and the subvoxel bisection and orientation estimation cannot compensate. This is more central than the orientation-estimation accuracy, which is at least quantitatively assessed on CT, and more specific than the general 'no ground-truth evaluation' critique. The proposed test—manual annotation on a held-out MRI subset with per-level analysis—directly measures whether the heuristic generalizes. A positive result (small, level-independent errors) would weaken the concern; a negative result would invalidate the 'subpixel-accurate' claim as stated. Because the reader's verdict is already CONDITIONAL and this concern is the same one, no verdict change is warranted; the condition should be that the raycasting directions are validated across levels and populations before the headline claim is accepted.","tokens_in":7398,"tokens_out":7472,"duration_ms":90333,"concrete_test":"Select 20 of the 37 MRI spines, deliberately including scoliotic cases, and have two experts independently annotate the transverse and spinous process tips on the sagittal T2w MRI. Run the released TPTBox code to generate the rule-based POIs. Compute per-level (C1–S1) Euclidean distances between the algorithm's POIs and the mean expert annotation. If the mean error exceeds one in-plane voxel in any level group (~1 mm in SI, ~3.3 mm in LR), or if errors systematically vary with vertebral level, the empirically chosen raycasting directions fail to generalize and the 'subpixel-accurate' claim is unsupported. If errors stay within voxel size and are level-independent, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 defines the transverse-process ray as a = 0.5·l + 0.5·p and the spinous-process ray as a = d + 0.2·p, stating these coefficients were 'empirically chosen to best match the observed anatomical trajectories'. These fixed coefficients are the only mechanism producing the process endpoints that later define muscle and ligament attachment sites in the multibody model. They are not derived from an anatomical model, not stratified by vertebral level, and not tested for sensitivity. Cervical, thoracic, and lumbar processes differ in orientation (e.g., thoracic spinous processes slope steeply downward; cervical transverse processes are short and often bifid), and scoliotic deformities further change local geometry. The only quantitative evaluation (Table 1) measures posterior-direction orientation on CT, not downstream POI position on MRI. The qualitative 37-spine expert check can detect gross failures but cannot reliably detect systematic 1–3 mm shifts of process tips. Because the pipeline is deterministic and segmentation-based, an incorrect ray direction produces a reproducible, systematic error that the subvoxel bisection cannot correct. Thus the central claim of reliable subpixel-accurate POI extraction from MRI rests on an unvalidated heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper adapts a rule-based CT key-point extraction pipeline to MRI for generating points of interest (POIs) for multibody spine models. The method takes SPINEPS segmentations from sagittal T2-weighted MRI, estimates a local vertebral coordinate system using a 2D projection of posterior structures, computes process endpoints by raycasting along hand-specified direction vectors, and localizes vertebral body corners and ligamentum flavum points via subvoxel bisection. The authors evaluate the orientation estimate on 90 CT vertebrae (Table 1) and report a qualitative expert assessment on 37 full-spine segmentations, arguing that the method is robust to low resolution and spinal curvature. The central claim is subpixel-accurate key-point extraction from MRI for biomechanical digital twins.","tokens_in":7771,"tokens_out":3461,"duration_ms":43334,"significance":"If fully validated, the contribution is valuable: it extends an existing CT-based biomechanical modeling pipeline to MRI, which is radiation-free and better suited for soft-tissue characterization; it is open-source and builds on publicly available segmentation tools; and it directly addresses vertebral orientation in the presence of deformities such as scoliosis. The orientation estimation results (mean error 1.72±1.76°, 80% below 3°, 100% below 10°) are promising. However, the headline claim of 'subpixel-accurate key-point extraction' is not yet supported by the evidence: the only quantitative evaluation concerns vertebral orientation, not key-point position, and the downstream MBS validation is anecdotal. The hand-picked raycasting directions are a potential source of systematic, reproducible error that the current validation cannot detect.","major_comments":[{"comment":"The raycasting directions for the transverse and spinous processes (a = 0.5l + 0.5p and a = d + 0.2p) are stated to be 'empirically chosen to best match the observed anatomical trajectories,' but no sensitivity analysis or independent validation of these coefficients is provided. These endpoints define muscle and ligament attachment sites in the MBS, so a systematic error here directly propagates into the biomechanical model. The coefficients are not stratified by vertebral level, despite known differences between cervical, thoracic, and lumbar process orientations, nor tested on pathological spines. Because the pipeline is deterministic, an incorrect direction produces a reproducible bias that the subvoxel bisection cannot correct. I recommend a quantitative comparison of the computed process endpoints against expert-annotated landmarks or CT-derived POIs on the same subjects, with per-","section":"Section 2.2"},{"comment":"The claim of 'subpixel-accurate' or 'subvoxel-accurate' landmark placement conflates numerical precision with measurement accuracy. The bisection search with interpolation guarantees that the reported coordinate can be placed at subvoxel resolution, but it does not establish that this coordinate is anatomically correct. Given the low left–right resolution of clinical sagittal T2-weighted MRI (3–4 mm), the segmentation boundary itself is uncertain, and no ground-truth comparison of key-point positions on MRI is presented. The phrase 'subpixel-accurate' in the Abstract should be replaced by a more precise statement, such as 'subvoxel-precise extraction,' unless a direct accuracy evaluation against independent reference landmarks is added.","section":"Section 2.3 and Abstract"},{"comment":"The validation for the final key-point output is anecdotal. The authors state that two experts assessed 37 full-spine segmentations and that failures occurred only when the segmentation was flawed, but no quantitative scoring, inter-observer variability, error metrics, or case-level statistics are reported. The statement that 'we validated our point placement using an existing MBS framework [13]' is also weak evidence, because the framework originates from the same group and no quantitative comparison of MBS outputs (e.g., joint angles, loads, or ligament forces) is given. This section does not substantiate the reliability of the POIs for biomechanical simulation. A quantitative landmark error study, or at minimum a structured expert rating with defined criteria and agreement statistics, is needed.","section":"Section 3.2"},{"comment":"The orientation evaluation is the only quantitative experiment and it is limited in important ways: it is performed on CT, not on the target MRI modality; the manual angular measurements lack inter-observer variability; and the 90 vertebrae appear to come from non-pathological or mixed VerSe2020 subjects, with no explicit inclusion of scoliotic cases. The paper's motivation emphasizes scoliosis, but the only scoliosis evidence is the qualitative Figure 4. While the orientation method is segmentation-based and may transfer across modalities, the reported numbers do not directly support the MRI-specific claim. Please clarify how representative the evaluated vertebrae are of the target population and report measurement repeatability.","section":"Section 3.1 and Table 1"}],"minor_comments":[{"comment":"Heading typo: 'V ertebra Orientation' should be 'Vertebra Orientation'.","section":"Section 2.1"},{"comment":"Heading typo: 'Sub-V oxel' should be 'Sub-Voxel'.","section":"Section 2.3"},{"comment":"The terms 'subpixel' and 'subvoxel' are used interchangeably; please standardize, as 'pixel' and 'voxel' have different meanings in 2D and 3D.","section":"Abstract and Section 2.3"},{"comment":"Please report the acquisition parameters of the 37 MRI scans, the number of subjects, and any inclusion/exclusion criteria, as this is important for reproducibility and for judging generalizability.","section":"Section 3.2"},{"comment":"The predefined precision threshold for the bisection search is not stated. Please give the actual value used in the experiments.","section":"Section 2.3"},{"comment":"The table would benefit from confidence intervals or a statistical test between the three methods; as presented, the standard deviations are large and the reader cannot assess the significance of the improvement.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically well-structured and the open-source release is a clear strength. My main concern is that the validation does not yet match the strength of the claims, particularly the 'subpixel-accurate key-point extraction' headline. I believe this is fixable within the scope of a revision by adding a quantitative landmark accuracy study and tightening the language. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful, incremental adaptation of the known CT-based POI extraction to low-resolution MRI, with a genuinely nice 2D projection trick for vertebral orientation. The code is released, the pipeline is deterministic and fast. But the paper's central claim of subpixel-accurate extraction is not actually measured, and the process endpoints in the multibody model depend on manually chosen raycasting coefficients that are never tested for sensitivity or across vertebral levels. That is the soft spot to probe.\n\nWhat's new: the orientation estimation (project posterior structures onto a plane orthogonal to the up/down axis, then take the 2D center of mass) is simple and seems effective. Their evaluation on 90 CT vertebrae shows mean error of 1.72 degrees, 80% under 3 degrees, and no catastrophic failures above 10 degrees — a real improvement over naive 3D CMS. The subvoxel bisection for the corpus corners is a sensible way to handle anisotropic resolution, and the open-source TPTBox implementation is a plus for reproducibility. The paper is honest about the limitations of rule-based systems and about the frontal-ligament ambiguity in the simulation.\n\nSoft spots, in order: (1) 'Subpixel-accurate' is asserted from the bisection procedure, not validated against any ground truth for landmark positions. The expert check on 37 spines can catch gross failures, not 1–3 mm shifts. (2) The raycasting directions in Section 2.2 — a=0.5l+0.5p for transverse processes and a=d+0.2p for spinous process — are admitted to be empirically chosen. These coefficients determine the very points that feed the MBS muscle and ligament attachments. No ablation, no stratification by cervical/thoracic/lumbar level, no analysis of how errors there affect simulation outputs. Given the known anatomical variation (downward-sloping thoracic spinous processes, bifid cervical transverse processes), this is a real gap. (3) The orientation validation is on CT, not MRI, and uses a single manual measurement per vertebra with no inter-observer variability. Since the target is MRI, the claimed robustness to low-resolution scans is only partially supported. (4) The downstream validation uses the authors' own MBS framework and is qualitative: agreement between two same-lab tools is weaker evidence than an independent benchmark.\n\nNone of these are deal-breakers. The central approach is sound and the failures are addressable: a small landmark evaluation study (even on a handful of MRI scans against a second rater) and a sensitivity sweep of the raycasting coefficients would substantially strengthen the claims. The paper is honest, well-written, and the code availability is a real asset.\n\nWho is this for? Anyone building spine biomechanical models from MRI, especially scoliosis-oriented work, and groups wanting a deterministic alternative to learned landmark detection. It deserves a serious referee; with revisions, it could be a solid methods paper.","headline":"Useful incremental MRI adaptation of the authors' CT landmark pipeline with a clever orientation fix; the 'subpixel-accurate' claim is not measured and the hand-picked raycasting directions are a real gap.","tokens_in":8229,"tokens_out":4127,"would_cite":true,"duration_ms":47839,"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":"MRI-only landmark extraction for spinal digital twins achieves subpixel accuracy by estimating each vertebra's orientation instead of assuming image alignment.","keywords":["digital twin","spine biomechanics","MRI landmark extraction","multibody simulation","vertebral orientation","subvoxel bisection","T2-weighted MRI","rule-based point extraction"],"falsifier":"Scan a cohort of patients with both CT and sagittal T2 MRI, compute POIs from the CT pipeline and from this MRI pipeline in the same world coordinate system, and measure the Euclidean distance between matched landmarks; systematic deviations that grow with scoliotic curvature or with transverse-process morphology would show the empirical raycasting directions do not generalize. A simpler check: rotate a high-resolution volume by known angles and verify that the computed landmarks rotate with the vertebra rather than drifting relative to it.","tokens_in":7349,"feed_emoji":"🦴","tokens_out":5656,"duration_ms":60985,"temperature":0.7,"pith_summary":"This paper extends a CT-based rule-based landmark extraction method so it works on sagittal T2-weighted MRI, where left-right resolution is only 3–4 mm and vertebrae may be rotated relative to the image axes. The authors claim that by estimating each vertebra's local orientation from the segmentation and using raycasting with subvoxel bisection, the pipeline places the points needed for multibody spinal models—vertebral corners, process endpoints, ligamentum flavum insertions—accurately enough for biomechanical simulation. This matters because MRI is radiation-free and resolves soft tissues, making subject-specific spinal digital twins feasible for large studies and for populations where CT is not justified. The orientation estimation is the quantified improvement: their 2D projection method reaches a mean angular error of 1.72 degrees, below 3 degrees in 80% of tested vertebrae, whereas a naive center-of-mass approach averaged 5.78 degrees.","feed_headline":"Spinal digital twins get subpixel MRI landmarks","feed_subtitle":"Rule-based pipeline estimates vertebral orientation and key points from low-resolution T2 MRI, no CT needed.","key_machinery":"The load-bearing mechanism is a per-vertebra orthonormal local coordinate system combined with raycasting and sub-voxel bisection. The up/down axis comes from a spline through vertebral body centers of mass; the front/back axis comes from projecting arcus and spinous process masks onto the plane orthogonal to that spline and connecting their 2D center of mass to the corpus center; left/right is the cross product. All landmarks are then defined by raycasting along local axes, or empirically chosen combinations of them, with corpus corners and ligamentum flavum points located by a 2D bisection search that interpolates at subvoxel positions.","core_discovery":"The paper's central claim is that a segmentation-only, rule-based pipeline can extract subpixel-accurate key points from low-resolution T2-weighted MRI by replacing image-axis assumptions with a per-vertebra local coordinate system and by locating boundaries through interpolated bisection. The front/back axis is computed by projecting the arcus and spinous process masks onto the plane orthogonal to the spline-defined up/down axis and joining their 2D center of mass to the vertebral corpus center; left/right follows as the cross product. Process endpoints are then obtained by raycasting along local directions that were empirically chosen to match anatomy, while vertebral body corners and liga","pith_inferences":["The fixed raycasting coefficients (0.5 lateral + 0.5 posterior for transverse processes, 0.2 posterior offset for the spinous process) are anatomical heuristics, not derived quantities; a matched CT-MRI comparison in the same subjects would reveal whether they generalize to scoliotic or morphologically unusual spines.","Since the pipeline's accuracy ceiling is set by the segmentation, improving the effective resolution of the left-right axis, for example through super-resolution or better segmentation, should directly improve landmark placement.","The large intervertebral forces at corner points that the authors attribute to the ligament model, not the landmarks, suggest MRI's soft-tissue contrast could eventually support extracting the actual ligament path instead of heuristic attachment offsets."],"forward_implications":["Multibody simulations of the spine can be built from routine MRI alone, avoiding radiation exposure and adding soft-tissue context that CT-based models lack.","Rotated or scoliotic spines no longer need manual alignment before landmark extraction, since the pipeline derives a local coordinate system for every vertebra.","Because the method runs on segmentation masks and completes in under a minute per spine, it scales to large cohorts and can feed subject-specific simulation studies.","The landmark outputs can be corrected and used as training data for learning-based predictors, which the paper identifies as the natural next step beyond accumulated rule exceptions."],"supporting_citations":[{"why":"Supplies the original CT-based point-of-interest extraction framework and the multibody simulation setup used to validate point placement.","marker":"[13]"},{"why":"Provides the SPINEPS segmentation model that produces the vertebra and subregion masks the pipeline operates on.","marker":"[15]"},{"why":"Companion SPINEPS reference used for automated whole-spine segmentation from T2-weighted MRI.","marker":"[9]"},{"why":"Supplies the 90 vertebrae from a public CT dataset used to quantify the orientation estimation error against manual measurements.","marker":"[18]"}],"fun_headline_variants":["MRI-only spine landmarks hit subpixel accuracy","Rule-based MRI pipeline marks spine key points","Low-res MRI yields precise spine landmarks, no CT","Spine digital twins gain subpixel MRI precision","CT-free spine modeling gets exact MRI key points"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the hand-picked raycasting directions—0.5 lateral plus 0.5 posterior for transverse processes and down plus 0.2 posterior for the spinous process—match real anatomy across patients; if those direction coefficients are wrong for unseen populations or pathologies, every process endpoint shifts systematically even when the orientation estimate is correct.","fun_headline_variants_meta":{"raw":{"variants":["MRI-only spine landmarks hit subpixel accuracy","Rule-based MRI pipeline marks spine key points","Low-res MRI yields precise spine landmarks, no CT","Spine digital twins gain subpixel MRI precision","CT-free spine modeling gets exact MRI key points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000398,"raw_usage":{"total_tokens":1884,"prompt_tokens":674,"completion_tokens":1210,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":418,"completion_tokens_details":{"reasoning_tokens":1138}},"tokens_in":418,"tokens_out":1210,"duration_ms":13749,"temperature":1.0,"reasoning_tokens":1138,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:18:20.399461+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Scan a cohort of patients with both CT and sagittal T2 MRI, compute POIs from the CT pipeline and from this MRI pipeline in the same world coordinate system, and measure the Euclidean distance between matched landmarks; systematic deviations that grow with scoliotic curvature or with transverse-process morphology would show the empirical raycasting directions do not generalize. A simpler check: rotate a high-resolution volume by known angles and verify that the computed landmarks rotate with the vertebra rather than drifting relative to it.","supporting_citations":[{"cited_title":"Frontiers in bioengineering and biotechnology10, 862804 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the original CT-based point-of-interest extraction framework and the multibody simulation setup used to validate point placement."},{"cited_title":"https://doi.org/10.1007/s00330-024-11155-y, https://doi.org/10","cited_arxiv_id":null,"evidence_quote":"Provides the SPINEPS segmentation model that produces the vertebra and subregion masks the pipeline operates on."},{"cited_title":"European Radiology Experimental7(1), 70 (2023)","cited_arxiv_id":null,"evidence_quote":"Companion SPINEPS reference used for automated whole-spine segmentation from T2-weighted MRI."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 90 vertebrae from a public CT dataset used to quantify the orientation estimation error against manual measurements."}],"review_version":1}