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

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Biomedical Image Analysis · Depth: Expert, extended

Summary

A new decoupled framework addresses the pervasive problem of single-mask annotation noise in vascular computed tomography (CT) datasets. This method leverages "cross-sectional patch self-consistency," observing that anatomically similar patches along vessel centrelines should have consistent masks. By sampling cross-sectional patches, retrieving intensity-equivalent neighbors via scalable vector search, and computing a patch-level noise score from statistical mask disagreement, the framework produces interpretable and auditable noise evidence. Experiments on the ImageCAS coronary CT dataset, comprising 1000 scans and approximately 3x10^6 patches, show that quality-weighted training improves CPR-DSC by 1.4% and reduces HD-95 by 4.1%. The analysis also reveals systematic annotation biases, with transverse and oblique vessels exhibiting a 5.1x higher error rate than axis-aligned structures, and correlations to cross-sectional area and intensity. Precomputation takes about 6 hours on an NVIDIA RTX 4090.

Key takeaway

For Machine Learning Engineers developing segmentation models for vascular CT, this framework offers a critical tool for dataset quality control. You can audit single-mask annotations directly, identifying and mitigating noise without costly multi-rater fusion or training-coupled methods. Incorporating quality maps into your training loss can improve boundary-sensitive metrics like CPR-DSC and HD-95, leading to more accurate and robust models, especially for challenging oblique or small vessels.

Key insights

Cross-sectional patch self-consistency detects single-mask annotation noise in tubular anatomy.

Principles

• Similar image patches imply consistent masks.
• Mask disagreement signals unreliable annotation.

Method

Sample Bishop-frame cross-sectional patches, retrieve intensity-equivalent neighbors via vector search, then compute a patch-level noise score from mask disagreement to generate scan-level quality maps.

In practice

• Improve segmentation robustness via quality-weighted training.
• Audit annotation quality with explicit patch-pair evidence.
• Identify systematic annotation biases.

Topics

• Annotation Noise Detection
• Vascular Segmentation
• Computed Tomography
• Self-Consistency Principle
• Quality-Weighted Training
• ImageCAS Dataset

Code references

• zhuyinheng/hidden_noise

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Decoupled Single-Mask Annotation Noise Detection via Cross-Sectional Patch Self-Consistency — Cross-sectional patch self-consistency detects single-mask annotation noise in tubular anatomy, providing auditable evidence.
• Attention-Based Prototype Calibration for Multi-Rater Few-Shot Medical Image Segmentation — Prototype calibration with attention effectively models multi-rater variability in few-shot medical image segmentation.
• Rethinking Noise-Robust Training for Frozen Vision Foundation Models: A Cross-Dataset Benchmark with a Case Study of Small-Loss Failure — No universal noisy-label learning method exists for frozen Vision Foundation Models; selection must be regime-aware.
• Learning a Sampling-Free Variational DNN Plugin from Tiny Training Sets to Refine OOD Segmentation With Uncertainty Estimation — VarDeepPCA refines OOD medical image segmentations by learning anatomical geometric priors from small in-distribution datasets using sampling-free variational learning.
• Patch Knowledge Transfer for Efficient AI-Generated Image Quality Assessment — Patch Knowledge Transfer efficiently assesses AI-generated image quality by distilling multi-level knowledge from a complex teacher to a simpler student model.
• Noise-Aware Framework for Correcting Corrupted Labels — CANOLA corrects corrupted labels by estimating noise and iteratively refining them for robust model training.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.