PaperlayerView exact revision on arXiv

Source-linked AI summary

Hard Sample Aware Noise Robust Learning for Histopathology Image Classification

Chuang Zhu, Wenkai Chen, Ting Peng, Ying Wang, Mulan Jin

arXiv:2112.03694v1eess.IVcs.AIcs.CVcs.LGq-bio.QM

TL;DR

Noisy labels are difficult to avoid in histopathology datasets, and existing approaches may fail to distinguish informative hard samples from harmful noise. The paper combines prediction-history-based EHN detection and iterative label correction with NSHE; it reports improved performance on real-world clinical data and noise tolerance across tested noise ratios.

  • Problem

    Noisy labels are common in manual annotation, and distinguishing informative hard samples from harmful noisy ones remains challenging.

    Source: §I · pp. 1–2

  • Method

    The method uses training prediction histories for EHN detection, iteratively corrects labels through self-training, and applies NSHE to the resulting almost-clean data.

    Source: §I · p. 2, B. Label Correction · p. 4, C. Noise Suppressing and Hard… · p. 5

  • Results

    The method reports superior performance on clinical pathology data and tolerates about 22%, 34%, 40%, 44%, and 49% noise at 0%, 10%, 20%, 30%, and 40% noise ratios, respectively.

    Source: §I · p. 2, E. Ablation Study · p. 10

  • Takeaways & Limitations

    The approach is reported to save more hard clean samples and can be applied to real-world noisy data without clean annotations.

    Source: §I · p. 2

Abstract

from arXiv · show

Deep learning-based histopathology image classification is a key technique to help physicians in improving the accuracy and promptness of cancer diagnosis. However, the noisy labels are often inevitable in the complex manual annotation process, and thus mislead the training of the classification model. In this work, we introduce a novel hard sample aware noise robust learning method for histopathology image classification. To distinguish the informative hard samples from the harmful noisy ones, we build an easy/hard/noisy (EHN) detection model by using the sample training history. Then we integrate the EHN into a self-training architecture to lower the noise rate through gradually label correction. With the obtained almost clean dataset, we further propose a noise suppressing and hard enhancing (NSHE) scheme to train the noise robust model. Compared with the previous works, our method can save more clean samples and can be directly applied to the real-world noisy dataset scenario without using a clean subset. Experimental results demonstrate that the proposed scheme outperforms the current state-of-the-art methods in both the synthetic and real-world noisy datasets. The source code and data are available at https://github.com/bupt-ai-cz/HSA-NRL/.

I. INTRODUCTION

Histopathology classification can aid cancer diagnosis, but clean labeled datasets are difficult to obtain and annotation is vulnerable to variability and workload. The paper addresses the difficulty of distinguishing informative hard samples from noisy ones with a two-phase learning method.

I. INTRODUCTION

  • Clean labels are expensive to collect, and manual annotation can be noisy because it requires expertise, varies across observers, and is time-consuming [7],.Existing histopathology classification work often assumes clean training data.Source: §I · p. 1
  • Hard samples can improve training, while overfitting to noisy labels degrades performance; distinguishing hard clean samples from noise remains unresolved –.The cited histopathology methods do not adequately separate informative hard samples from harmful mislabeled ones –.Source: §I · p. 2
  • The proposed two-phase method uses sample prediction history for easy/hard/noisy detection, then self-training for label correction and NSHE to suppress noise and enhance hard samples.The EHN detector is integrated into self-learning, and NSHE operates on the corrected data.Source: §I · p. 2
  • The method is reported to save more clean samples and to achieve superior performance on clinical pathology data from a Beijing hospital without requiring clean annotations.The authors also state that it can be applied to real-world datasets without clean annotations.Source: §I · p. 2

II. RELATED WORKS

Prior label-correction approaches may depend on clean samples or involve all pseudo-labels in correction training, where misleading samples can impair the correction model. Learning-to-teach methods select samples for training but may overlook informative hard samples that also have large losses.

A. Noisy Label Correction

  • Some noisy-label correction methods require clean annotations, while self-learning alternatives use pseudo-labels but face uncertainty about which labels are correct –.The cited approaches include replacing labels with pseudo-labels and self-learning with multi-prototypes.Source: A. Noisy Label Correction · p. 2
  • Using all pseudo-labels to train a correction model risks including misleading samples; EHN uses prediction history to identify likely noisy samples for removal.The authors state that this removal improves correction quality and helps preserve hard samples for NSHE.Source: A. Noisy Label Correction · p. 2

B. Learning to Teach

  • Co-teaching and related methods select small-loss samples, but hard and noisy samples both have large losses, so informative hard samples may be ignored [19], [20].The paper responds with pre-discarding of likely noisy samples and NSHE to enhance hard samples while suppressing remaining noise.Source: B. Learning to Teach · p. 3

III. METHOD

The method first corrects labels and filters likely noise to form an almost-clean dataset, then uses NSHE to train a robust classifier. EHN distinguishes easy, hard, and noisy samples from prediction histories and supports iterative self-training label correction.

A. Architecture

  • The method has two phases: label correction produces an almost-clean dataset, which NSHE uses to train the final robust classifier.The architecture trains two models in NSHE, with M2 selecting training data for M1.Source: A. Architecture · p. 3, Figure 1 · p. 3

B. Label Correction

  • EHN uses mean prediction probabilities over training history to select easy samples; the remaining clean samples below the threshold are treated as hard.The paper notes that clean samples mostly have higher mean prediction probabilities than noisy samples, but some clean samples fall below the threshold.Source: B. Label Correction · p. 3
  • EHN trains an MLP on prediction histories from easy samples with added label noise, then classifies the remaining data as hard or noisy.The added-noise ratio matches the original dataset's estimated noise ratio; EHN's output is used for subsequent correction and filtering.Source: B. Label Correction · p. 4
  • The correction model trains on easy and hard samples, then supplies pseudo-labels for hard and noisy samples during iterative self-training.Hard samples are included because the EHN classifier's predictions are not treated as fully reliable.Source: B. Label Correction · p. 4
  • Post-processing removes samples judged noisy using agreement between original and pseudo-labels, retaining the resulting almost-clean dataset for NSHE.The procedure removes noisy-set samples whose labels were not changed and hard-set samples whose labels were changed.Source: B. Label Correction · p. 4

C. Noise Suppressing and Hard Enhancing (NSHE)

NSHE takes the almost-clean dataset and trains a noise-robust classifier with two same-initialized models. It combines low-confidence sample filtering, a momentum-updated model, and focal loss to suppress noisy samples while emphasizing hard ones.

C. Noise Suppressing and Hard Enhancing (NSHE)

A. Dataset

The evaluation uses five datasets, including three histopathology datasets and two natural-image datasets. DigestPath2019 and Camelyon16 receive synthetic noise, while Chaoyang and Webvision represent real-world noisy-label settings.

A. Dataset

  • DigestPath2019 and Camelyon16 are patch datasets with separate training and test source images or slides; both receive synthetic label noise at 10%, 20%, 30%, and 40%.DigestPath2019 uses 256 × 256 patches; Camelyon16 is preprocessed similarly.Source: A. Dataset · pp. 5–6
  • Fig. 5 presents selected sample patches from DigestPath2019, Camelyon16, and Chaoyang.Its panels correspond to those three datasets in that order.Source: A. Dataset · pp. 5–6, Figure 5 · p. 6
  • Chaoyang is a real-world noisy dataset: three pathologists labeled patches, consensus cases formed the test set, and about 40% of training cases had inconsistent opinions.For inconsistent training cases, one of the three opinions was selected at random.Source: A. Dataset · p. 6
  • CIFAR-10 uses symmetric and asymmetric synthetic noise, while Webvision uses the first 50 Google-image classes and contains real-world noisy labels.CIFAR-10 noise ratios were 20%, 50%, and 80% for symmetric noise and 40% for asymmetric noise; Webvision has 65,944 training and 2,500 test samples.Source: A. Dataset · p. 6

B. Implementation and Parameter Settings

Implementation settings differ between medical and natural-image datasets, and the evaluation reports standard classification metrics. For multiclass tasks, per-class metrics are macro-averaged.

B. Implementation and Parameter Settings

  • Medical datasets use ResNet-34 with Adam, momentum 0.9, batch size 96, and learning rate 0.001; label correction runs 30 epochs and NSHE runs 40.The learning rate is linearly reduced after epoch 15.Source: B. Implementation and Parameter Settings · p. 6
  • CIFAR-10 uses ResNet-32 with SGD for 300 epochs per phase, while Webvision uses Inception-ResNet v2 with SGD for 80 epochs per phase.Their learning rates and batch sizes are 0.02 and 128 for CIFAR-10, and 0.01 and 32 for Webvision.Source: B. Implementation and Parameter Settings · p. 6
  • The easy-sample ratio is 0.1 for 80% noise and 1 − 1.5 × ρ otherwise; the discard ratio is 0.1 × ρ.For Chaoyang and Webvision, ρ is estimated using the noise cross-validation algorithm [20].Source: B. Implementation and Parameter Settings · p. 6

C. Evaluation Criteria

  • Evaluation uses accuracy, precision, recall, F1, AUC, and ROC curves; ACC is calculated from true and false positives and negatives.ROC plots false-positive rate against true-positive rate, and AUC is the area under that curve.Source: C. Evaluation Criteria · pp. 6–7
  • For multiclass tasks, precision, recall, F1, ROC, and AUC are computed per class and macro-averaged.Tables I and II report average test metrics over three runs with standard deviations for DigestPath2019 and Camelyon16.Source: C. Evaluation Criteria · p. 7, Table I · p. 7, Table II · p. 7

D. Objective Comparison

The method is compared with general noisy-label methods and medical-data methods, using the same network architecture. It almost outperforms state-of-the-art methods across medical noise ratios and outperforms them on the reported natural-image datasets.

Compared methods

  • The comparison uses the same network architecture and includes general noisy-data baselines and methods developed for medical data.The authors report using open-source implementations for methods 2–6 and re-implementing methods 1 and 7.Source: D. Objective Comparison · pp. 7–8

Medical datasets

  • Across DigestPath2019 and Camelyon16 noise ratios of 10%–40%, the method almost always outperforms state-of-the-art methods; on Chaoyang it leads on every criterion.The reported criteria are test ACC, AUC, F1 Score, Precision, and Recall.Source: Table III · p. 7, D. Objective Comparison · p. 8

Natural computer vision datasets

E. Ablation Study

Ablations indicate that EHN detection contributes the largest performance gain, followed by NSHE, while label correction also helps. The authors report that one self-training round is sufficient and that the method tolerates substantially more noise than a simple ResNet-34 at matched performance.

Component ablations

  • Removing NSHE lowers test accuracy by an average of about 1.5%, indicating the contribution of its hard-sample-enhancing training stage.The ablation trains a single model on the label-corrected dataset without NSHE.Source: E. Ablation Study · p. 9
  • Removing EHN detection and NSHE lowers test accuracy by an average of about 2.8%; EHN detection gives the largest component gain, followed by NSHE.All components provide a gain at every noise ratio.Source: E. Ablation Study · p. 9
  • Removing the whole label-correction phase lowers test accuracy by an average of about 1.2%.The ablation trains on the original data; the learning rate is adjusted to match the epoch count.Source: E. Ablation Study · p. 9

Rounds and noise tolerance

  • 22%, 34%, 40%, 44%, and 49% noise are tolerated at matched performance to a simple ResNet-34 trained with 0%, 10%, 20%, 30%, and 40% noise, respectively.More self-training rounds consume more computing resources but bring little gain, so the experiments use one round.Source: E. Ablation Study · p. 10, Figure 10 · p. 10, Figure 9 · p. 10

F. Analysis of EHN Detection Scheme

The EHN detector’s accuracy remains effective across tested noise ratios, with performance relatively stable beyond a training-epoch threshold and under reasonable changes to the easy-sample ratio. Prediction-probability distributions on the artificial dataset help explain why its training history can distinguish sample types.

Effectiveness and convergence

Parameter analysis

Why Da works

  • Clean and noisy samples trained on artificial Da have mean prediction-probability distributions similar to those in original D, supporting EHN’s use of Da.Some clean samples on Da cannot be distinguished from noisy ones by mean prediction probability alone, and some easy samples become hard on Da.Source: F. Analysis of EHN Detection Scheme · p. 11, Figure 15 · p. 11

G. Hard and Noisy Sample Analysis

Hard and noisy samples show different training-history patterns, while some noisy samples in another class’s center behave as hard-to-predict samples late in training. The label-correction phase removes more noise than a baseline at the same retained-sample count while protecting hard samples.

Training behavior

  • In early training, hard samples tend to have more learning events and noisy samples more forgetting events; the pattern reverses late, while hard samples have more events overall.A learning event crosses the labeled-class prediction probability from below 0.5 to above 0.5; forgetting crosses it in the opposite direction.Source: G. Hard and Noisy Sample Analysis · p. 12, Figure 17 · p. 12

Gradient behavior

  • Hard samples tend to have higher adjacent-epoch prediction-probability changes than noisy samples.The authors suggest some noisy samples centered in another class are “super hard” to fit, with their optimization effect suppressed by surrounding clean samples.Source: G. Hard and Noisy Sample Analysis · p. 12

Sample visualization

Label-correction benefit

V. CONCLUSION

The proposed method uses training prediction history to distinguish hard samples from noisy ones, preserve more hard clean samples, and improve noise-robust histopathology classification.

V. CONCLUSION

  • Training prediction history distinguishes hard samples from noisy samples, allowing more hard clean samples to be saved during noise removal.The authors report that this supports improved model performance.Source: §V · p. 13
  • Under a co-learning architecture, NSHE uses different parameter updating speeds for the two models to further suppress noisy-sample interference.Source: §V · p. 13
  • The authors report compelling performance on noisy datasets and say the method can be applied directly to real-world noisy scenarios.Source: §V · p. 13
  • Future work will investigate hard-sample-aware semantic segmentation, including malignant-tissue segmentation.Source: §V · p. 13
Loading 2112.03694v1…