Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Noisy Label Classification using Label Noise Selection with Test-Time Augmentation Cross-Entropy and NoiseMix Learning

Hansang Lee, Haeil Lee, Helen Hong, Junmo Kim

TL;DR

The paper tackles the challenge of learning from datasets with noisy labels, focusing on medical imagery. It introduces a two-step approach: first, a Label Noise Selection stage using Test-Time Augmentation Cross-Entropy to accurately separate clean and noisy labels, and second, a classifier training stage with NoiseMix that blends clean and noisy data via MixUp-inspired mixing. Experiments on the ISIC-18 skin lesion dataset show that TTA cross-entropy improves noisy-label detection over standard cross-entropy and TTA uncertainty, while NoiseMix delivers superior classification performance and the greatest robustness to label noise compared with existing methods. The method achieves strong results without introducing bespoke loss functions or weighting schemes, suggesting practical utility for real-world noisy medical imaging tasks.

Abstract

As the size of the dataset used in deep learning tasks increases, the noisy label problem, which is a task of making deep learning robust to the incorrectly labeled data, has become an important task. In this paper, we propose a method of learning noisy label data using the label noise selection with test-time augmentation (TTA) cross-entropy and classifier learning with the NoiseMix method. In the label noise selection, we propose TTA cross-entropy by measuring the cross-entropy to predict the test-time augmented training data. In the classifier learning, we propose the NoiseMix method based on MixUp and BalancedMix methods by mixing the samples from the noisy and the clean label data. In experiments on the ISIC-18 public skin lesion diagnosis dataset, the proposed TTA cross-entropy outperformed the conventional cross-entropy and the TTA uncertainty in detecting label noise data in the label noise selection process. Moreover, the proposed NoiseMix not only outperformed the state-of-the-art methods in the classification performance but also showed the most robustness to the label noise in the classifier learning.

Noisy Label Classification using Label Noise Selection with Test-Time Augmentation Cross-Entropy and NoiseMix Learning

TL;DR

The paper tackles the challenge of learning from datasets with noisy labels, focusing on medical imagery. It introduces a two-step approach: first, a Label Noise Selection stage using Test-Time Augmentation Cross-Entropy to accurately separate clean and noisy labels, and second, a classifier training stage with NoiseMix that blends clean and noisy data via MixUp-inspired mixing. Experiments on the ISIC-18 skin lesion dataset show that TTA cross-entropy improves noisy-label detection over standard cross-entropy and TTA uncertainty, while NoiseMix delivers superior classification performance and the greatest robustness to label noise compared with existing methods. The method achieves strong results without introducing bespoke loss functions or weighting schemes, suggesting practical utility for real-world noisy medical imaging tasks.

Abstract

As the size of the dataset used in deep learning tasks increases, the noisy label problem, which is a task of making deep learning robust to the incorrectly labeled data, has become an important task. In this paper, we propose a method of learning noisy label data using the label noise selection with test-time augmentation (TTA) cross-entropy and classifier learning with the NoiseMix method. In the label noise selection, we propose TTA cross-entropy by measuring the cross-entropy to predict the test-time augmented training data. In the classifier learning, we propose the NoiseMix method based on MixUp and BalancedMix methods by mixing the samples from the noisy and the clean label data. In experiments on the ISIC-18 public skin lesion diagnosis dataset, the proposed TTA cross-entropy outperformed the conventional cross-entropy and the TTA uncertainty in detecting label noise data in the label noise selection process. Moreover, the proposed NoiseMix not only outperformed the state-of-the-art methods in the classification performance but also showed the most robustness to the label noise in the classifier learning.
Paper Structure (8 sections, 5 equations, 4 figures, 1 table)

This paper contains 8 sections, 5 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Label Noise Selection with Test-Time Augmentation Cross-Entropy
  4. Classifier Training with NoiseMix
  5. Experiments and Results
  6. Datasets and Experimental Details
  7. Results
  8. Conclusions

Figures (4)

  • Figure 1: A pipeline of the proposed method.
  • Figure 2: A pipeline of the proposed label noise selection method.
  • Figure 3: Histograms of various prediction scores for the label noise (orange) and the clean label (blue) data.
  • Figure 4: ROC curves of the cross-entropy (red), TTA uncertainty (blue), and the proposed TTA cross-entropy (black) for detection of the label noise data in the label noise selection process with different label noise ratios $r$.