ARNAI: Artifact Removal Network based on Autoencoding and Inpainting for Robust Spinal Image Segmentation and Measurement
Sang-Jin Park, Jinyoung Choi, Seokwon Kim, Seungeon Song, Insu Park, Dougho Park, Taeyeon Kim, Youjin Lee, et al.
cs.CV · cs.AI
Quick verdict
Claims ARNAI (VQ-GAN OOD suppression + Telea inpainting) cuts postop L4-L5 Cobb error ~70% internally; external significance unproven (q=0.077).
Summary
This retrospective two-center study proposes RSM (Restore, Segment, and Measure), a framework for automated spinopelvic measurement on postoperative lateral lumbar radiographs. Its core component, ARNAI, removes spinal-implant artifacts without any implant-labeled training data: a VQ-GAN trained exclusively on implant-free radiographs reconstructs each image, suppressing out-of-distribution implants; the pixel-wise difference localizes implants (gamma correction, Otsu thresholding, morphology), and Telea inpainting fills the erased regions. The restored image feeds an FCBFormer segmenter (L1–L5, sacrum, femoral heads) whose landmarks yield PT, SS, LL and L4–L5 segmental Cobb angle. On the internal test set, mean DSC rose from 0.8136 to 0.8697, with the largest gains at L3–L5; on 91 internal implant radiographs, L4–L5 Cobb angle error fell from 15.6–16.2° to 4.7° (≈70%), with ICC improving from 0.18 to 0.54–0.59 — significant after Benjamini–Hochberg correction (q≈10⁻⁸–10⁻⁹). On the external cohort the same effect was suggestive only (q=0.077), and RSM's external L4–L5 ICC was negative (−0.03/−0.04) despite a lower mean error. PT/LL/SS gains were modest and non-significant. Limitations include a single external center, only 88 of 217 external images analyzable, and no segmentation-level significance testing.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- ARNAI removes implant-related artifacts from postoperative radiographs using only implant-free training data, via VQ-GAN out-of-distribution suppression plus difference-map localization and Telea inpainting.
- Adding ARNAI to FCBFormer raises mean segmentation DSC from 0.8136 to 0.8697 (internal) and 0.8091 to 0.8521 (external), with the largest gains at L3–L5.
- On 91 internal implant-containing radiographs, RSM reduces mean L4–L5 segmental Cobb angle error from 15.6–16.2° to 4.7° (~70%), significantly after BH correction (q≈10⁻⁸–10⁻⁹), with ICC improving from 0.18 to 0.54–0.59.
- ICCs for PT, LL, and SS exceed 0.70 on implant-containing radiographs with RSM, though these parameters show no significant post-correction improvement over baseline.
Strengths
No implant-labeled training data required
ARNAI's VQ-GAN (§2.2) is trained exclusively on implant-free images; implants are suppressed as out-of-distribution content and localized by input-reconstruction differencing (gamma correction, Otsu, morphology), avoiding the scarce annotated implant datasets that prior work (refs 7, 11, 12) depended on.
Two-center design with held-out external test set
2,486 internal radiographs (1,683 patients, 2017–2024) split 8:1:1; 217 external postoperative radiographs from a separate tertiary hospital (443 screened, 226 excluded by predefined criteria) used only for testing, probing domain shift to lower-quality, non-standard postures (§2.1, Table 1 external rows).
Statistical pipeline is more careful than typical radiograph-AI papers
Large, targeted effect on the intended failure mode
Honest mechanism interpretation
Weaknesses
External L4–L5 SCA result is internally incoherent as reported
Benefit is essentially confined to one parameter
No uncertainty quantification for segmentation
Common-subset conditioning and external attrition
Reproducibility and baseline breadth
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method paper (arXiv 2609.07013) targets a real clinical gap: deep-learning spinopelvic measurement degrades on postoperative lateral lumbar radiographs because metal implants obscure anatomy, and prior fixes either exclude implant cases or require scarce annotated implant data. The authors propose RSM, whose preprocessing module ARNAI is trained **only** on implant-free images: a VQ-GAN reconstructs the radiograph, out-of-distribution implants are suppressed, and the input–reconstruction difference (gamma correction → Otsu → morphological cleanup) yields an implant mask that is erased and filled by Telea inpainting (§2.2). Downstream, FCBFormer segments L1–L5, sacrum, and femoral heads; landmarks feed PT, SS, LL, and L4–L5 segmental Cobb angle. The central claims are a DSC gain 0.8136→0.8697 (internal) and a ~70% reduction of L4–L5 SCA error on 91 internal implant radiographs (16.2°/15.6° → 4.7°), plus ICC improvement from 0.18 to 0.54–0.59.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does ARNAI need implant-containing images or implant masks for training?
No. Its VQ-GAN is trained exclusively on implant-free radiographs. Because implants are out-of-distribution, they vanish in the reconstruction; the pixel-wise difference between input and reconstruction localizes them (Otsu thresholding), and Telea inpainting fills the erased regions. Avoiding annotated implant data is the method's main practical advantage.
How much did ARNAI improve Cobb angle measurement?
On 91 internal postoperative radiographs with implants, the mean L4–L5 segmental Cobb angle error fell from 15.6–16.2° (two expert raters) to 4.7°, roughly 70%, and the ICC rose from 0.18 to 0.54–0.59. This was statistically significant after multiple-comparison correction (q≈10⁻⁸–10⁻⁹). The external-center version of this result was weaker: suggestive only (q=0.077).
Did the results replicate at the external hospital?
Partially. Segmentation DSC improved externally (0.8091→0.8521 mean), and mean L4–L5 Cobb error dropped (14.2°→9.5°), but the adjusted significance test missed (q=0.077) and the external ICC for that parameter was negative (−0.03/−0.04), meaning residual large-error cases remain. External validation is therefore incomplete rather than confirmatory.
Is code or data available for this paper?
The paper text contains no code or data availability statement, and both cohorts are private clinical archives. Independent replication would require institutional radiograph access and a reimplementation of the VQ-GAN differencing + Telea inpainting + FCBFormer pipeline; several key hyperparameters are only in the supplementary material.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Supplementary Materials A.1–A.5 and Supplementary Tables 1–10 are cited but not part of the provided text, so hyperparameters, Bland-Altman details, and ablation numbers could not be verified.
- Figures 1–4 are available only via captions in the text extraction; qualitative claims about restoration quality rest on the captions.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.