Decoupled Single-Mask Annotation Noise Detection via Cross-Sectional Patch Self-Consistency

This paper presents a decoupled method for detecting noise in single-mask vascular CT annotations without multirater fusion or dependence on network training. It samples cross-sectional patches along vessel centrelines, retrieves intensity-equivalent neighbours with scalable vector search, and computes patch-level noise scores from statistical mask disagreement. The output is explicit image-mask evidence for each flagged region, together with scan-level quality maps that can support dataset auditing or quality-weighted training. Experiments on a coronary CT dataset report that transverse and oblique vessels have error rates 5.1 times higher than axis-aligned structures, with additional associations involving cross-sectional area and intensity.