Source-linked AI summary
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, Donghoon Yang, Jaeman Cho, Joongwon Yang, Mansu Kim, Heumdai Kwon, Hong Gyu Baek, Dae Chul Cho, Injung Kim
TL;DR
Postoperative metallic artifacts limit automated spinopelvic assessment on radiographs, despite their routine use after lumbar fusion. RSM uses ARNAI to mitigate those artifacts before segmentation and measurement. It improved segmentation and spinopelvic measurement, with the largest reported benefit in L4–L5 SCA error.
Problem
Postoperative metallic artifacts undermine deep-learning assessment on radiographs, while existing mitigation has relied on image exclusion or scarce labeled data.
Source: §1 · p. 4Method
RSM uses ARNAI, trained on implant-free images, to localize and mitigate artifacts before segmentation and spinopelvic parameter measurement.
Source: §1 · p. 4Results
0.8697 versus 0.8136 was the mean DSC with ARNAI, and L4–L5 SCA error decreased by approximately 70%.
Source: §4 · p. 7Takeaways & Limitations
The findings support RSM's feasibility for more robust postoperative spinopelvic measurement under implant-related distortion.
Source: §4 · pp. 6–7The external cohort came from one relatively small hospital, reference evaluation included only adequate-landmark images, and experiments focused on L4–L5.
Source: §4 · p. 7
Abstract
from arXiv · showhide
Purpose: This study aims to develop an AI framework applicable for postoperative imaging for automated measurement of spinopelvic parameters on radiographs with robustness to the presence of spinal implants. Materials and Methods: We retrospectively reviewed lateral lumbar spine radiographs from two institutions (Internal: January 2017--December 2024; External: October 2021--September 2025). We developed the Restore, Segment, and Measure (RSM) framework, incorporating a novel Artifact Removal Network based on Autoencoding and Inpainting (ARNAI) to mitigate implant-related artifacts in postoperative radiographs. Segmentation and spinopelvic parameter (PT, LL, SS, SCA) measurement performance were assessed using Wilcoxon signed-rank tests and intraclass correlation coefficients. Results: When ARNAI was added to a recent Transformer-based segmentation model, FCBFormer, the mean DSC increased to 0.870 from 0.814, with marked gains at L3--L5 and smaller improvements at L1--L2. On 91 radiographs with implants, the mean L4--L5 segmental Cobb angle error decreased to 4.7 {\deg} from 15.6--16.2 {\deg}, an average error reduction of 70%. The ICC for L4--L5 segmental Cobb angle improved to 0.54 (Rater 1) and 0.59 (Rater 2) from 0.18, and ICCs for pelvic tilt, lumbar lordosis, and sacral slope all exceeded 0.70. The improvement in L4--L5 segmental Cobb angle error was statistically significant in the internal implant-containing cohort after correction for multiple comparisons. Conclusion: The proposed RSM framework improved automated spinopelvic parameter measurement in implant-containing postoperative radiographs. By mitigating implant-related artifacts, ARNAI improved segmentation and downstream measurement accuracy, with the greatest benefit observed for L4--L5 segmental Cobb angle estimation, where the mean error was reduced by approximately 70%.
1 Introduction
Postoperative implants can undermine automated spinopelvic assessment by obscuring anatomy on radiographs. The study introduces an artifact-removal framework designed to support robust measurement without implant-containing training data.
1 Introduction
- Metallic artifacts compromise the generalizability of deep-learning assessment in postoperative radiographs, which are routinely used to evaluate outcomes after lumbar fusion.Manual measurement is time-consuming and subject to interobserver variability.Source: §1 · p. 4
- Earlier mitigation strategies either excluded implant-containing images or relied on scarce annotated implant datasets or expert-labeled keypoints.Source: §1 · p. 4
- RSM uses ARNAI, trained exclusively on implant-free images, to localize and mitigate implant artifacts before segmentation and measurement of PT, SS, LL, and L4–L5 SCA.L4–L5 was selected for its clinical relevance, common involvement in degenerative lumbar pathology, and frequent inclusion in lumbar fusion surgery.Source: §1 · p. 4
2 Materials and Methods
The study used retrospective radiographs from two institutions and developed a two-stage pipeline that restores implant-affected images before segmentation and measurement. Performance was assessed against expert references and compared with a baseline pipeline.
2.1 Data Acquisition and Preparation
- The internal archive contributed 2,486 radiographs from 1,683 patients; external testing used 217 implant-containing radiographs from an independent hospital.Internal images were collected from January 2017 to December 2024; external images were collected from October 2021 to September 2025.Source: §2.1 · p. 4
2.2 Model Development
- ARNAI localizes implants by comparing an implant-free-trained VQ-GAN reconstruction with the input, then erases detected regions and applies Telea inpainting.The reconstruction is used only for localization because anatomical detail outside implant regions is not faithfully preserved; Telea was chosen over diffusion-based inpainting for better preservation of non-implant regions.Source: §2.2 · p. 5
- FCBFormer segments L1–L5, the sacrum, and femoral heads; computer-vision landmarks from those masks support PT, SS, LL, and L4–L5 Cobb-angle measurement.Evaluation compared segmentation masks with ground truth using DSC and compared parameter predictions with expert measurements; baseline and RSM errors were tested with paired Wilcoxon signed-rank tests and Benjamini–Hochberg correction.Source: §2.2 · p. 5, §2.3 · pp. 5–6
3 Results
Adding ARNAI improved lumbar segmentation, with gains most pronounced at implant-affected levels. In implant-containing radiographs, RSM most notably reduced L4–L5 SCA error, with a statistically significant internal-cohort improvement after multiple-comparison correction.
3.1 Segmentation Performance
- 0.8697 versus 0.8136 was the overall mean DSC on the internal test set with ARNAI; external gains were strongest at L4 (+0.1677) and L5 (+0.1448).Internal DSC gains were +0.096 at L3, +0.184 at L4, and +0.136 at L5; sacrum and F1 DSC decreased slightly.Source: §3.1 · p. 6
3.2 Spinopelvic Parameter Estimation
- 4.7° versus 16.2° (Rater 1) and 15.6° (Rater 2) was the mean L4–L5 SCA error in 91 implant-containing internal radiographs, an average reduction of approximately 70%.L4–L5 SCA ICC improved to 0.54 and 0.59 from 0.18; ICCs for PT, LL, and SS were generally above 0.70.Source: §3.2 · p. 6
- Inter-rater ICC(A,1) was 0.941–0.988 for global parameters and 0.871 for L4–L5 SCA; intra-rater ICC(3,1) was 0.956–0.999 across parameters.Repeated measurements showed minimal differences, indicating high repeatability and negligible measurement bias.Source: §3.2 · p. 6
- q = 1.8 × 10^-8 and 2.2 × 10^-9 were the corrected internal-cohort values for lower L4–L5 SCA error with RSM; external q = 0.077 exceeded the 0.05 threshold.PT, SS, and LL showed no significant internal-cohort differences after correction.Source: §3.2 · p. 6
4 Discussion
The discussion presents RSM as a feasible approach to postoperative measurement under implant-related distortion. It highlights improved segmentation and measurement while noting scope and generalizability limitations.
4 Discussion
- With ARNAI preprocessing, RSM increased mean DSC from 0.8136 to 0.8697 and reduced L4–L5 SCA error by approximately 70%, improving delineation of implant-obscured anatomical boundaries.The authors say these gains enhance robustness of anatomical landmark detection for automated measurement rather than merely improving image appearance.Source: §4 · p. 7
- ARNAI was trained on implant-free images and requires no implant-mask annotation, unlike prior approaches that excluded implant cases or required annotated implant data.Source: §4 · p. 7
- Local segmental measurements remain more sensitive to endplate definition and image quality than global spinopelvic parameters.The discussion notes that inter-rater agreement was relatively lower for L4–L5 SCA; it interprets ARNAI as stabilizing contour continuity and downstream landmark inference.Source: §4 · p. 7
- Evaluation focused on L4–L5 and an external cohort from one relatively small referral hospital, limiting conclusions about other surgical levels, constructs, and more challenging examinations.External reference evaluation included only radiographs with adequate visualization of key landmarks.Source: §4 · p. 7