[Paper Review] Learning to Segment from Noisy Annotations: A Spatial Correction Approach

Jiachen Yao, Yikai Zhang+4 more|arXiv (Cornell University)|Jul 21, 2023

Machine Learning and Data ClassificationComputer Science3 citations

TL;DR

This paper proposes a novel Markov-based label noise model for medical image segmentation that captures spatial correlation and bias in noisy annotations, and introduces a spatial correction (SC) method to progressively recover true labels using a small clean validation set. The approach outperforms state-of-the-art methods on both synthetic and real-world noisy segmentation benchmarks, with theoretical guarantees and empirical validation showing robustness even under high noise or minimal validation data.

ABSTRACT

Noisy labels can significantly affect the performance of deep neural networks (DNNs). In medical image segmentation tasks, annotations are error-prone due to the high demand in annotation time and in the annotators' expertise. Existing methods mostly assume noisy labels in different pixels are extit{i.i.d}. However, segmentation label noise usually has strong spatial correlation and has prominent bias in distribution. In this paper, we propose a novel Markov model for segmentation noisy annotations that encodes both spatial correlation and bias. Further, to mitigate such label noise, we propose a label correction method to recover true label progressively. We provide theoretical guarantees of the correctness of the proposed method. Experiments show that our approach outperforms current state-of-the-art methods on both synthetic and real-world noisy annotations.

Motivation & Objective

• To address the critical challenge of label noise in medical image segmentation, where annotations are error-prone due to high time and expertise demands.
• To model segmentation label noise not as i.i.d. but as spatially correlated, reflecting real-world annotation errors near object boundaries.
• To develop a label correction method that mitigates bias introduced by such spatially correlated noise using only a small clean validation set.
• To provide theoretical guarantees that minimal validation data suffice for full bias correction and label recovery.
• To demonstrate the effectiveness of the method on both synthetic and real-world noisy segmentation datasets, outperforming existing SOTA approaches.

Proposed method

• Proposes a Markov process model for segmentation label noise that simulates boundary distortions using two Bernoulli variables for expansion/shrinkage decisions and spatial-dependent strength along the boundary.
• Incorporates a sparse random flipping component to allow occasional mislabels in interior or exterior regions, capturing realistic annotation errors.
• Develops a label correction algorithm that estimates and removes bias using a clean validation set, with theoretical proof that even one validation sample suffices for full bias correction.
• Introduces an iterative variant of the correction method (SC) that alternates between training a segmentation model and refining labels, improving robustness to general noise patterns.
• Designs the method to be architecture-agnostic, allowing integration with any deep neural network backbone for segmentation.
• Employs a spatially aware noise model that generates realistic noisy masks—continuous distortions of true boundaries—contrasting with unrealistic i.i.d. noise patterns.

Experimental results

Research questions

• RQ1 Can a Markov process model effectively capture the spatial correlation and bias inherent in real-world segmentation label noise?
• RQ2 Is it possible to correct label noise in segmentation using only a small number of clean validation annotations, with theoretical guarantees of correctness?
• RQ3 How does the proposed spatial correction method compare to existing SOTA methods in terms of performance under high or realistic noise levels?
• RQ4 Does the iterative refinement of labels through model training and correction lead to consistent performance gains across diverse datasets?
• RQ5 Can the method generalize to real-world noisy annotations beyond synthetic noise settings?

Key findings

• On the JSRT dataset with high noise levels (S_E and S_S settings), the proposed SC method achieves a mean DSC of 91.51±0.62, significantly outperforming the next best method (CLE at 86.85±0.84).
• On the LIDC-IDRI dataset with real-world noise, SC achieves a mean DSC of 53.72±1.70, surpassing the best baseline (CDR at 49.41±0.79).
• In ablation studies, SC maintains high performance even when noise levels are increased to extreme levels, while baselines like GCE and ELR degrade rapidly.
• SC remains effective with as few as one clean validation sample, achieving a DSC of 82.30±0.10 on Cityscapes, outperforming all baselines including CLE (74.41±0.12).
• The iterative version of SC converges and consistently improves segmentation performance across multiple benchmarks, demonstrating robustness to complex noise patterns.
• Qualitative results in Figure 5 show that SC effectively corrects boundary distortions and restores true segmentation masks, even in cases with severe noise.