ARNAI: Artifact Removal Network Based on Autoencoding and Inpainting for Robust Spinal Image Segmentation and Measurement

Abstract

Purpose To develop an artificial intelligence framework, robust to spinal implants, for automated measurement of spinopelvic parameters on postoperative radiographs. Materials and Methods Lateral lumbar spine radiographs from two institutions (internal dataset [ n = 2486]: January 2017-December 2024; external dataset [ n = 217]: October 2021-September 2025) were retrospectively reviewed. The Restore, Segment, and Measure (RSM) framework was developed, incorporating a novel Artifact Removal Network based on Autoencoding and Inpainting (ARNAI) to mitigate implant-related artifacts. Segmentation and spinopelvic parameter (including segmental Cobb angle [SCA]) measurement performance relative to expert reference measurements were assessed using the Dice similarity coefficient (DSC), mean absolute error (MAE), Wilcoxon signed-rank tests, and Benjamini-Hochberg correction for multiple comparisons. Results With ARNAI added to the segmentation pipeline, the mean DSC increased to 0.870 from 0.814, with marked gains at L3-L5. For implant-containing radiographs in the internal test set, rejected radiographs decreased by 64.21% (95 to 34), and the L4-L5 SCA MAE decreased to 4.7° from 16.2°-15.6° (~ 70% reduction); this remained significant after correction. For the external test set, rejected radiographs decreased by 65.45% (110 to 38), and the L4-L5 SCA MAE decreased to 9.5°-9.7° from 14.2°-14.5° (~ 33% reduction); however, this improvement did not remain significant after correction, and agreement with expert references remained limited. Conclusion The RSM framework improved automated spinopelvic parameter measurement in implant-containing postoperative radiographs, with the greatest benefit observed for L4-L5 SCA estimation in the internal dataset. In the external dataset, error was reduced but agreement remained limited. © RSNA, 2026