{"id":"d07bdc6d-f6db-4a31-89b2-d8e490a346c7","arxiv_id":"2411.09204","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A 3D U-Net trained on CT scans can partially predict a synthetic missing ribcage segment treated as an implant, but accuracy is too low for clinical use.","lead":"Using a 3D U-Net on CT scans, the authors train a model to predict a missing ribcage region that they treat as an implant, reporting low Dice scores around 0.25. The paper is an early feasibility probe for automating ribcage implant design, a task traditionally done by manual CAD, but the current results are far from clinical use.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never tests implant generation: its ground truth is the patient's own bone inside a synthetic 64×64×64 cuboid (Sec. 2.1), so the reported DSC/HD measure bone inpainting, not surgical implant design; the central claim is therefore unsupported.","rationale":"The central claim is that the model generates ribcage implants from CT scans. For that claim to be true, the target variable in the training/evaluation protocol must be an implant. Section 2.1 defines the ground truth as Ig = Sg·(1−Md), where Sg is the thresholded bone volume and Md is a fixed 64×64×64 cuboid defect mask placed between relative heights 0.5 and 0.75 to include the sternum. Thus Ig is the original bone occupying the synthetic cuboid. Perfect prediction would reconstruct the patient's own bone, not design an implant: real implants require extensions/fixation, must accommodate resected/pathological tissue, and are judged by biomechanical and surgical criteria, not by voxel overlap with bone that may no longer exist. Consequently Table 1's DSC/HD numbers, including the best DSC of 0.2524, characterize a synthetic inpainting task. The paper's 'preliminary' framing is honest, but the title, abstract, and 'first deep-learning approach for automated ribcage implant generation' claim overstate what is demonstrated. I see no internal contradiction that rescues this: the reported numbers are poor, and even an oracle solution to this protocol would not be a clinically usable implant. The reader's REJECT verdict is therefore appropriate. I would add that Fig. 3's caption, saying the ground-truth Ig is input to the network, is an independent methodological red flag; if literal, it would make the evaluation circular, though the manuscript gives no implementation to resolve it. The absence of code, error bars, and a clear data split further underscores that the feasibility conclusion rests on an unvalidated proxy rather than on demonstrated implant-generation capability.","tokens_in":5501,"tokens_out":7351,"duration_ms":82132,"concrete_test":"Obtain the code/checkpoints and rerun the exact protocol while withholding the ground-truth Ig from the network input at test time, then evaluate on a clinically realistic defect benchmark: defects derived from real rib resections with surgeon-designed implants (including fixation margins) as ground truth. If the model cannot be applied to non-cuboid defects or its DSC drops far below the reported 0.2524, the central claim of implant generation fails; if it performs comparably on realistic defects, the concern is resolved.","verdict_should_be":"REJECT","load_bearing_attack":"The claim that this is an automated ribcage-implant generator requires that the training target in Sec. 2.1 be a plausible implant. It is not. The defect mask Md is a fixed 64×64×64 zero cuboid inserted between heights 0.5 and 0.75, and the ground-truth implant is defined as Ig = Sg·(1−Md), i.e., the patient's own thresholded bone that originally occupied that cuboid. Clinically, a ribcage implant is not the patient's original bone: after tumor resection or trauma the missing tissue is often pathological or absent, and the designed implant includes fixation margins, screw holes, and biomechanical properties aimed at restoring chest-wall function. Optimizing voxel overlap with a pre-existing bone segment inside an arbitrarily placed synthetic hole does not test the ability to design such an implant. Thus the headline experimental result (DSC 0.2524) is a measurement of inpainting fidelity on an artificial missing-bone task, not of implant generation. The paper's 'preliminary' wording is honest, but the title, abstract, and contribution list make the broader claim. A separate red flag strengthens the concern: Fig. 3's caption says both Rd and Ig are 'processed and input into the network,' which, if literal, would be label leakage; at minimum the manuscript does not specify the actual input at inference, so the evaluation protocol is unclear.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RibCageImp, a 3D U-Net-based framework that takes CT scans as input and is claimed to generate patient-specific 3D ribcage implants. The method creates synthetic defects by placing a fixed 64x64x64 zero mask between relative heights 0.5 and 0.75 of the volume, and defines the ground-truth implant Ig = Sg * (1 - Md), i.e., the patient's own thresholded bone inside that synthetic hole. The network is trained with combinations of Dice, MSE, extra-region-removal, and gap-filling losses, and evaluated on the RibFrac dataset with DSC and Hausdorff distance. The reported best DSC is 0.2524 with HD 148.90 mm, while the text also claims a DSC of 0.52 in unspecified 'specific scenarios'. The paper concludes that the results demonstrate feasibility of automated ribcage implant generation.","tokens_in":5814,"tokens_out":6072,"duration_ms":70959,"significance":"The idea of automating ribcage implant design is clinically relevant, and the paper honestly describes its results as preliminary. The use of real CT data from RibFrac and the systematic comparison of loss configurations are useful starting points. However, the central claim that this is a deep-learning framework for ribcage implant generation is not supported by the experimental design. The ground-truth 'implant' is not an independent clinical target but the patient's own bone within an arbitrarily placed synthetic defect; the evaluation therefore measures synthetic-defect inpainting, not implant design. The quantitative results are low, the dataset split is inconsistent, no baselines or error bars are provided, and the inference protocol is unclear. As a preliminary inpainting study the paper could be a modest contribution, but as a contribution to implant generation it does not meet the standard for a serious journal publication.","major_comments":[{"comment":"The training and evaluation target is defined as Ig = Sg * (1 - Md), where Md is a fixed 64x64x64 zero cuboid placed at relative heights 0.5-0.75. This makes the ground truth the patient's own thresholded bone occupying an artificially introduced hole, not a clinical implant. A surgical implant for a resected or fractured ribcage is not the original bone: the resected tissue may be pathological or absent, and the implant must include fixation margins, material properties, and biomechanical considerations. Consequently, the optimization in Eq. (1) and the metrics in Table 1 measure the fidelity of bone inpainting on a self-generated label, and the central contribution claim (Section 1, first bullet) is not supported by the experiments as designed.","section":"Sec. 2.1, Eq. (1)"},{"comment":"The best quantitative result reported in Table 1 is DSC 0.2524 and HD 148.90 mm, yet the text states that the model achieves 'a DSC of 0.52 in specific implant generation scenarios'. The conditions defining these scenarios are not specified, and the number is inconsistent with the table. The evaluation also lacks confidence intervals, multiple random seeds, or any statistical significance testing, so the claim of demonstrated feasibility is not substantiated.","section":"Sec. 3.1, Table 1"},{"comment":"The dataset split is described inconsistently: Section 2.1 states that RibFrac provides 420 training, 160 test, and 80 validation samples, while Section 3 states that experiments use 300 training samples and 160 test cases. This discrepancy must be resolved because it affects the reproducibility and interpretation of all reported metrics.","section":"Sec. 2.1 vs Sec. 3"},{"comment":"The caption of Figure 3 states that both the defective region Rd and the ground truth implant Ig are 'processed and input into the network'. If Ig is used as an input at inference, or even as an auxiliary input during training, this constitutes label leakage, and the evaluation would be invalid. The manuscript never specifies the exact input tensor at inference time; it must clarify that Ig is used only as the training target and state what exactly is fed to the network when a prediction is made.","section":"Fig. 3, Sec. 2.2"},{"comment":"The network architecture is described as '3D U-Net architecture and an EfficientNet-b0 encoder', but EfficientNet-b0 is a 2D backbone pretrained on ImageNet. The paper does not explain how this encoder is adapted to volumetric 3D input, what the encoder's role is in a 3D U-Net, or which parts of the network are 3D convolutions. This is a reproducibility-critical ambiguity in the method.","section":"Sec. 2.2"}],"minor_comments":[{"comment":"The minimization in Eq. (1) is written as 'min over Rd', but the defective ribcage Rd is an input, not an optimization variable; this should be minimized over the network parameters or over Ip.","section":"Eq. (1)"},{"comment":"The notation I_g^{-1} is confusing; if the intended operation is the complement or element-wise inversion of the binary mask, it should be written explicitly (for example, 1 - Ig) to avoid ambiguity with a matrix inverse.","section":"Eqs. (4)-(7)"},{"comment":"The caption uses the phrase 'complete ribcage with predicted ground truth Rd + Ig', which is contradictory; this should be 'complete ribcage with predicted implant Rd + Ip' or similar.","section":"Fig. 1"},{"comment":"Please provide details of the preprocessing pipeline: CT windowing, voxel spacing/resampling, and how the fixed input size of 256x256x128 is obtained, since these choices directly affect the geometry of the ribcage and the implant target.","section":"Sec. 2.2"},{"comment":"The Hausdorff distance should specify whether it is the maximum or the 95th percentile; the maximum Hausdorff distance is highly sensitive to outliers, and HD of 148.90 mm for a thoracic implant is not interpretable without this information.","section":"Sec. 3.1"}],"recommendation":"reject","confidential_remarks":"The manuscript has a genuine core problem: the labels are self-generated by cutting out a fixed cuboid from the patient's own bone, so the task is synthetic-defect inpainting, not implant design. This is not fixable by a local revision; it requires either a fundamental reframing as an inpainting study or access to real defect/implant data with clinical ground truth. Even as an inpainting study, the missing baselines, inconsistent data split, unclear inference protocol, and low quantitative results would need substantial additional work. I see no path to acceptance in a serious journal without changing the scope of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: this paper does not test implant generation. The ground truth implant Ig is defined as Sg*(1 - Md), where Md is a fixed 64x64x64 zero mask placed between heights 0.5 and 0.75. So the target is the patient's own thresholded bone occupying that cuboid. The network is learning to fill a synthetic hole with the original bone segment. That's bone inpainting, not surgical implant design. The title and abstract claim otherwise, and that mismatch is the core problem.\n\nWhat's actually new: it's the first deep-learning attempt at ribcage implant generation that I know of, and the authors are upfront that results are preliminary. The 3D U-Net with EfficientNet-B0 encoder and the four-term loss (Dice, MSE, ERR, GF) is a reasonable setup. Using the public RibFrac dataset is a plus. The qualitative figures honestly show extraneous regions and gaps.\n\nSoft spots beyond the central one. The dataset split is inconsistent: Section 2.1 says 420 training samples, Section 3 says 300. There's a claim of DSC 0.52 in 'specific scenarios' with no definition or table entry; the best table value is 0.2524 with HD 148.90 mm. No error bars, no baselines, no code. And the Figure 3 caption says both Rd and Ig are 'processed and input into the network,' which, taken literally, is label leakage. At minimum the inference-time input is undefined.\n\nNone of this is fatal to the possibility that a deep network could help with rib reconstruction, but as it stands the experiments don't support the stated contribution. The paper is honest about being preliminary, but that doesn't fix the label definition. I'd suggest the authors reframe it as synthetic defect inpainting, add a simple baseline (e.g., morphological fill), and report variance across the test set.\n\nWho's this for? A reader working on thoracic segmentation or implant CAD might glance at it, but the mismatch means the numbers aren't informative for implant design. I wouldn't want it in the archive without revision. My recommendation is reject in this form.","headline":"The paper overclaims implant generation when it actually trains on the patient's own bone in a synthetic hole, so the headline result is inpainting, not implant design.","tokens_in":6333,"tokens_out":3060,"would_cite":false,"duration_ms":34083,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a 3D U-Net trained on CT scans can produce a ribcage implant for a defective region and that this is the first deep-learning framework for automated ribcage implant generation.","keywords":["ribcage implant generation","3D U-Net","CT scan","deep learning reconstruction","thoracic implant","RibFrac dataset","Dice score","Hausdorff distance"],"falsifier":"Compare the trained network's output against a surgeon-approved ribcage implant or post-operative CT from a real patient who received such an implant; if the predicted implant does not match the real prosthesis's shape or does not align with the resected margins, the central claim of automated clinical implant generation fails.","tokens_in":5304,"feed_emoji":"🦴","tokens_out":10572,"duration_ms":102040,"temperature":0.7,"pith_summary":"This paper is trying to establish that automated ribcage implant generation from CT scans is feasible with deep learning, and that its 3D U-Net-based framework is the first to attempt it. The motivation is that manual, CAD-based implant design is slow and variable, while a learning-based pipeline could produce patient-specific implants directly from imaging. To test this, the authors simulate a defect by masking a $64 \\times 64 \\times 64$ cuboid in the rib/sternum region of CT volumes from the RibFrac dataset, and use the patient's own bone in that region as the ground-truth implant. Their best model, trained with a combination of MSE, extra-region-removal, and gap-filling losses, reaches a Dice similarity coefficient of 0.2524 and a Hausdorff distance of 148.90 mm on the test set, which they present as moderate but evidence of feasibility and a baseline for future work.","feed_headline":"First deep-learning pipeline maps CT to 3D ribcage implants","feed_subtitle":"A 3D U-Net fills a simulated missing rib segment; best Dice score is 0.2524, a feasibility signal.","key_machinery":"The central object is the ground-truth implant mask $I_g = S_g \\cdot (1 - M_d)$, obtained by thresholding the CT into bone $S_g$ and erasing a $64 \\times 64 \\times 64$ cuboid defect mask $M_d$ placed between relative heights 0.5 and 0.75 so that it includes the sternum. The prediction target is $I_p$, and the loss is $L_{rib} = L_{MSE} + L_{ERR} + L_{GF}$; the ERR term penalizes predicted voxels outside $I_g$ via elementwise multiplication by the inverse mask $I_g^{-1}$, while the GF term penalizes missing voxels through $I_p^{-1} \\circ I_g$. The 3D U-Net with an EfficientNet-b0 encoder carries the learning, and the loss combination is what pushes the model from over- and under-prediction toward the ground-truth region.","core_discovery":"The paper's central discovery claim is that a 3D U-Net with an EfficientNet-b0 encoder can learn a mapping from a defective ribcage volume $R_d$ to a predicted implant $I_p$ such that $R_d + I_p$ approximates the complete ribcage. The authors formalize the problem as minimizing a loss $L(I_g, I_p)$ over the implant prediction, and they show that combining MSE with two anatomy-aware terms -- an extra-region-removal loss that penalizes predicted voxels outside the ground truth and a gap-filling loss that penalizes missing voxels inside it -- improves Dice score from 0.1615 to 0.2524 and reduces Hausdorff distance from 220.94 mm to 148.90 mm compared with MSE alone. They interpret the still-low overlap as reflecting the genuine difficulty of ribcage geometry, patient variation, and local anatomical constraints, and they conclude that deep-learning ribcage implant generation is feasible but requires more data, better architectures, and further loss design.","pith_inferences":["The paper's evaluation stops at geometric overlap on a synthetic defect; it never measures whether a predicted implant would fit surgically, and a natural next test is to compare generated shapes against surgeon-approved CAD models or post-operative CTs of real rib reconstructions.","Because the ground truth implant is the patient's own bone, the model is effectively trained to regrow the patient's original anatomy, not to design a prosthesis with flanges, screw holes, or material thickness, and real implants may require exactly those non-anatomical features that the current loss functions would penalize.","A cheap falsification probe would be to replace the cuboid defect with irregular, clinically shaped resections; if performance collapses, the cuboid proxy, rather than the network, is the main reason the current results look plausible."],"forward_implications":["If the proposed loss combination (MSE + ERR + GF) is as useful as reported, other implant-generation or shape-completion tasks can adopt it to suppress extraneous predictions and fill gaps.","The results imply that a deep network can at least locate a missing ribcage region and produce a coarse shape for it, which is a necessary first step toward reducing manual CAD work.","Since the best configuration improved Dice score from 0.1615 with MSE alone to 0.2524 with the added losses, the additional loss terms are doing real work and are a productive direction for refinement.","The work supports the feasibility claim that CT scans contain enough information for a learning-based model to attempt ribcage reconstruction, even if current accuracy is moderate."],"supporting_citations":[{"why":"Supplies the RibFrac CT dataset used for training, validation, and testing in all experiments.","marker":"[12]"},{"why":"Provides the 3D U-Net architecture that processes the volumetric CT data and generates the implant prediction.","marker":"[5]"},{"why":"Establishes the prior deep-learning cranial implant work that this paper extends to the more complex ribcage anatomy.","marker":"[9]"},{"why":"Provides the Dice loss and Dice similarity coefficient used in one training configuration and in quantitative evaluation.","marker":"[13]"},{"why":"Supplies the Hausdorff distance metric used to assess surface accuracy of the predicted implants.","marker":"[14]"}],"fun_headline_variants":["Deep learning generates 3D ribcage implants from CT scans","First AI framework for 3D ribcage implant generation","3D U-Net maps ribcage defects to implants, Dice 0.2524","Automated ribcage implant design using deep learning","CT scans to 3D ribcage implants: a feasibility study"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a synthetic cuboid defect, with the ground-truth implant defined as the patient's own bone in that region, is an adequate proxy for a real ribcage implant; if actual implants require non-anatomical geometry, fixation margins, or material properties, the experiment never tests the paper's clinical claim.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning generates 3D ribcage implants from CT scans","First AI framework for 3D ribcage implant generation","3D U-Net maps ribcage defects to implants, Dice 0.2524","Automated ribcage implant design using deep learning","CT scans to 3D ribcage implants: a feasibility study"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2630,"prompt_tokens":905,"completion_tokens":1725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":1631}},"tokens_in":521,"tokens_out":1725,"duration_ms":106576,"temperature":1.0,"reasoning_tokens":1631,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:54:27.278012+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the trained network's output against a surgeon-approved ribcage implant or post-operative CT from a real patient who received such an implant; if the predicted implant does not match the real prosthesis's shape or does not align with the resected margins, the central claim of automated clinical implant generation fails.","supporting_citations":[{"cited_title":"nnu-net based segmentation and 3d reconstruction of uterine fi- broids with mri images for hifu surgery planning,","cited_arxiv_id":null,"evidence_quote":"Supplies the RibFrac CT dataset used for training, validation, and testing in all experiments."},{"cited_title":"Func- tional chest wall reconstruction with a biomechanical three-dimensionally printed implant,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D U-Net architecture that processes the volumetric CT data and generates the implant prediction."},{"cited_title":"3d u-net: learn- ing dense volumetric segmentation from sparse anno- tation,","cited_arxiv_id":null,"evidence_quote":"Establishes the prior deep-learning cranial implant work that this paper extends to the more complex ribcage anatomy."},{"cited_title":"Deep learning-based framework for auto- matic cranial defect reconstruction and implant model- ing,","cited_arxiv_id":null,"evidence_quote":"Provides the Dice loss and Dice similarity coefficient used in one training configuration and in quantitative evaluation."},{"cited_title":"Reconstruction of large chest wall defects using three-dimensional custom-made implant technology,","cited_arxiv_id":null,"evidence_quote":"Supplies the Hausdorff distance metric used to assess surface accuracy of the predicted implants."}],"review_version":1}