Source-linked AI summary
Robust Learning at Noisy Labeled Medical Images: Applied to Skin Lesion Classification
Cheng Xue, Qi Dou, Xueying Shi, Hao Chen, Pheng Ann Heng
TL;DR
Noisy labels are a significant challenge in medical image classification, where annotations require expertise and hard cases can resemble mislabeled examples. The paper combines online uncertainty sample mining with individual reweighting, and reports large accuracy improvements on noisy skin-lesion data.
Problem
Medical image classification depends on expert annotations, while deep networks can fit noisy labels and hard cases can complicate their identification.
Source: §1 · p. 1Method
An iterative framework combines online uncertainty sample mining with individual sample reweighting to suppress likely noisy labels while preserving hard examples.
Source: §1 · p. 2, §2.2 · p. 2Results
The authors report that accuracy improves by a large margin on noisy skin-lesion data.
Source: §4 · p. 4Takeaways & Limitations
The approach is end-to-end and requires neither prior estimation of the noise distribution nor extra clean data, and the authors state it can apply to other classification contexts.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Deep neural networks (DNNs) have achieved great success in a wide variety of medical image analysis tasks. However, these achievements indispensably rely on the accurately-annotated datasets. If with the noisy-labeled images, the training procedure will immediately encounter difficulties, leading to a suboptimal classifier. This problem is even more crucial in the medical field, given that the annotation quality requires great expertise. In this paper, we propose an effective iterative learning framework for noisy-labeled medical image classification, to combat the lacking of high quality annotated medical data. Specifically, an online uncertainty sample mining method is proposed to eliminate the disturbance from noisy-labeled images. Next, we design a sample re-weighting strategy to preserve the usefulness of correctly-labeled hard samples. Our proposed method is validated on skin lesion classification task, and achieved very promising results.
1. INTRODUCTION
Noisy labels are difficult to avoid in medical images, where annotation requires substantial expertise and deep networks can fit incorrect labels. Similar-looking hard cases further complicate distinguishing mislabeled examples from useful ones.
1. INTRODUCTION
- Medical images require expert annotation, and benign and malignant melanoma cases can look similar, complicating the identification of wrong labels.The introduction reports clinical melanoma dermoscopy diagnosis accuracy of 50% to 82.3%.Source: §1 · p. 1
- The authors focus on noisy labels, stating that class imbalance can be handled during preprocessing or data collection.Source: §1 · p. 1
- Prior approaches relying on noise-distribution assumptions or extra expert-labeled images face practical constraints, while treating weak-agreement samples as noise can degrade performance on hard medical images.The cited discussion also notes that hard samples and minority classes obstruct a noise-distribution assumption in melanoma data.Source: §1 · pp. 1–2
- The proposed iterative strategy combines online uncertainty mining to suppress noisy samples with individual reweighting to preserve hard samples and minority classes.The authors report that both components contribute to performance gains.Source: §1 · p. 2
2. METHOD
The framework uses a ResNet-101 backbone with online uncertainty-based sample mining and deep-feature-based reweighting. Mining excludes selected high-loss samples from backpropagation, while reweighting reduces the influence of likely noisy samples and preserves hard or minority-class examples.
2. METHOD
- The framework combines a ResNet-101 backbone, online uncertainty-based sample mining, and deep-feature-based sample reweighting.Source: §2 · p. 2
2.1. Online Uncertainty Sample Mining
- For each mini-batch, the method ranks samples by training loss and sets the losses of the top K high-loss samples to zero, leaving N-K samples to contribute gradients.It selects low-loss samples for backpropagation as training proceeds.Source: §2.1 · p. 2
2.2. Individual Re-weighting Loss
- The reweighting module uses 1-pLOF as each sample’s weight, updating weights every five epochs after an initial weight of 1.The combined objective includes online uncertainty-mining loss, weighted cross-entropy, and weight decay; the two trade-off parameters are set to 0.5.Source: §2.2 · pp. 2–3
3. EXPERIMENTS AND RESULTS
The experiments simulate label corruption in ISIC 2017, examine how networks fit noisy labels, and report accuracy gains from uncertainty mining and sample reweighting.
3.1. Data and Preprocessing
- The ISIC 2017 setup uses 3,582 training images and symmetrically corrupts top-loss samples within each class at γ=5%, 10%, 20%, or 40%.The corrupted label substitutes a different class label with probability γ.Source: §3.1 · p. 3
3.2. Memorization of Neural Networks to Noisy Data
- DNNs fit noisy training labels to zero error; clean samples fit earlier, while validation accuracy falls as training proceeds.All sample weights are set to 1 for the first five epochs.Source: §3.2 · p. 3
3.3. Quantitative Evaluation and Comparison
- OUSM improves accuracy across corrupted datasets, and individual reweighting adds further gains to the best Accuracy score; the method significantly outperforms.The comparison is reported in Table 1.Source: §3.3 · pp. 3–4, Table 1 · p. 4
4. CONCLUSION
The paper presents an iterative framework for noisy-labeled medical images and reports improved accuracy on noisy skin-lesion data. It also states that the method needs neither extra clean data nor prior noise-distribution estimation.
4. CONCLUSION
- The framework combines iterative online uncertainty sample selection with individual sample reweighting for noisy-labeled medical images.Source: §4 · p. 4
- On noisy skin-lesion data, the method improves accuracy by a large margin.Source: §4 · p. 4
- The method trains end-to-end without pre-estimating the noise distribution or using extra clean data, and the authors state it can apply in any classification context.Source: §4 · p. 4