Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Self-Relaxed Joint Training: Sample Selection for Severity Estimation with Ordinal Noisy Labels

Shumpei Takezaki, Kiyohito Tanaka, Seiichi Uchida

TL;DR

A new framework for training with “or-dinal” noisy labels that outperforms various state-of-the-art methods in experiments using two endoscopic ulcerative colitis datasets and a retinal Diabetic Retinopathy dataset.

Abstract

Severity level estimation is a crucial task in medical image diagnosis. However, accurately assigning severity class labels to individual images is very costly and challenging. Consequently, the attached labels tend to be noisy. In this paper, we propose a new framework for training with ``ordinal'' noisy labels. Since severity levels have an ordinal relationship, we can leverage this to train a classifier while mitigating the negative effects of noisy labels. Our framework uses two techniques: clean sample selection and dual-network architecture. A technical highlight of our approach is the use of soft labels derived from noisy hard labels. By appropriately using the soft and hard labels in the two techniques, we achieve more accurate sample selection and robust network training. The proposed method outperforms various state-of-the-art methods in experiments using two endoscopic ulcerative colitis (UC) datasets and a retinal Diabetic Retinopathy (DR) dataset. Our codes are available at https://github.com/shumpei-takezaki/Self-Relaxed-Joint-Training.

Self-Relaxed Joint Training: Sample Selection for Severity Estimation with Ordinal Noisy Labels

TL;DR

A new framework for training with “or-dinal” noisy labels that outperforms various state-of-the-art methods in experiments using two endoscopic ulcerative colitis datasets and a retinal Diabetic Retinopathy dataset.

Abstract

Severity level estimation is a crucial task in medical image diagnosis. However, accurately assigning severity class labels to individual images is very costly and challenging. Consequently, the attached labels tend to be noisy. In this paper, we propose a new framework for training with ``ordinal'' noisy labels. Since severity levels have an ordinal relationship, we can leverage this to train a classifier while mitigating the negative effects of noisy labels. Our framework uses two techniques: clean sample selection and dual-network architecture. A technical highlight of our approach is the use of soft labels derived from noisy hard labels. By appropriately using the soft and hard labels in the two techniques, we achieve more accurate sample selection and robust network training. The proposed method outperforms various state-of-the-art methods in experiments using two endoscopic ulcerative colitis (UC) datasets and a retinal Diabetic Retinopathy (DR) dataset. Our codes are available at https://github.com/shumpei-takezaki/Self-Relaxed-Joint-Training.

Paper Structure

This paper contains 20 sections, 8 equations, 5 figures, 7 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Training a classifier with noisy labels
  4. Ordinal regression with noisy labels
  5. Label smoothing for classification task
  6. Self-Relaxed Joint Training for Ordinal Noisy Labels
  7. Overview
  8. Problem formulation
  9. Soft labels from hard labels
  10. Sample selection by hard labels
  11. Updating the network parameters by soft labels
  12. How to apply our framework to other methods
  13. Experimental Evaluation
  14. Experimental Setup
  15. Quantitative Evaluation Results
  16. ...and 5 more sections

Figures (5)

  • Figure 1: (a) Traditional joint-training framework with dual-network model with sample selection for noisy labels. (b) The proposed self-relaxed joint training framework for learning with ordinal noisy labels. "M0’' stands for "Mayo 0."
  • Figure 2: Test accuracy curves. The width of the shading indicates the standard deviation in cross-validation.
  • Figure 3: Label precision curves. The blue and red curves show the label precisions by "Co-teaching" and "Co-teaching + Ours," respectively. The pink horizontal line shows $(1-\epsilon)$.
  • Figure 4: Test accuracy curves. The width of the shading indicates the standard deviation in cross-validation.
  • Figure 5: Label precision curves. The blue and red curves show the label precisions by "Co-teaching" and "Co-teaching + Ours," respectively. The pink horizontal line shows $(1-\epsilon)$.